Opencode vibe codes and releases like 3 projects a week at this points. Cool, just know its throwaway software to them. They seem really reckless when it comes to security. Opencode V2 basically threw security out the window. Theyve completely lost the plot as an org.
Their CEO spends all day on twitter talking about how much better he is than everyone. And how nobody can use agents like he can. Its quite obnoxious
Also just use CDP, way too many browser solutions for agents when chrome ships with the ability to control the browser at a low level.
If you want to wrap it in a jsonrpc api for mcp, this is like 4 hrs or a work for an llm, point them at a pkg thats uses CDP like go-rod.
They're totally right to not bother with security inside a harness, its just a terrible approach that creates permission fatigue and bad UX. Claude asking me to approve a 50 line arcane bash command doesn't help at all. Pi also has the same philosophy of not enforcing security within the harness.
Security should be enforced outside the harness with a sandbox. You can define what's allowed and what's not in one place and have strong enforcement guarantees instead of ad-hoc permission requests every minute.
> Security should be enforced outside the harness with a sandbox. You can define what's allowed and what's not in one place and have strong enforcement guarantees instead of ad-hoc permission requests every minute.
I agree in principle, but in practice most sandboxing technology (like bwrap) is too static for this use case: You grant access once during startup and can't extend/revoke without a restart. That's inconvenient when you realize halfway through your session that you need additional permissions, and it's insufficient for permissions that are only briefly needed (for the duration of one tool call).
A harness that controls the sandbox (e.g. for each individual tool invocation) can manage this much more effectively.
I almost never have to extend permissions of my sandbox. By default it can read/write the project tree, and read most files on the machine with specific secret locations excluded. If I do have to expand permissions and restart, I just resume the session.
A harness that controls the sandbox is exactly what I’m after. Tools don’t always need to share the same space, and I can isolate work much better if separate toolchains are isolated in sandboxes. My biggest gripe with the major harnesses is how they sandbox tool calls.
I have my agent request permissions for things like a Cloudflare API key (or I guide it or just come up with the request myself). These go in a simple queue, with justifications.
I review them and then have another agent with far more permissions create the tokens, with expiry and so on, and then hand the API tokens back.
The original agent doesn’t even run as my main user; it sits on a computer in a restricted account (easy to do with either macOS or Linux, or even Windows these days).
It has taken some rather long prompts to get it to self-report on itself when it tries to workaround.
You really wanting handle this at the sanbox layer? I have 100 different containers configurations, only a few differnet harness configurations.
You're just being stubborn, the best UX is through the harness. Most users dont even know how to configure a sandbox properly and dont even use one.
Ive known quite a few naive people that have gotten their home directories wiped and really gotten burned. Many users of these products dont even know how to use a terminal and these harness companies are marketing to the average joe.
I have a sandbox script and a simple wrapper that runs any command in that sandbox thus that "pi" -> without sandbox, "pisb" -> pi in the sandbox. I practically only use pisb.
This separation doesn't mean it must result in bad UX. It just needs somebody to make a nice UI/TUI to configure the sandbox and when running that it does its setup and open whatever command you want.
The problem with the harness doing sandboxing is that from my experience the simply break all the time leading to users ignoring properly configuring it and using abominations like codex' `Approve for me` where IT ASKS A SEPARATE AI and not the user.
My main issue is these are being marketed to average joes. You wont see anything on these harnesses documentation, opencode et al, who say it should be used in a sandbox. The people making them, opencode et al, scream everyday about how you're holding yourself back by using permissions without a single mention of sandbox.
Hundreds of thousands of people are using them naively, if they shipped with common sense strictish permissions instead of none at all, like what opencode opted to do it would be safer for everyone. Its much easier for a normal non programmer to configure a simple config file than it is for them to learn how to properly setup a container/vm, they will just not do the latter 9/10.
Honestly, a harness should always install heavily gaurded, remember when Grok code/build was instantly indexing your home directory and sending it back to HQ on the first request it made..
Like I wrote above I generally agree that the sandboxing should by enabled "by default" and should not be an extra opt-in that you have to know about first.
That being said I still believe this should be separated and composable at an application level and power users should be able to switch out the sandbox. This way you don't have to trust the harness since it doesn't know anything about the sandbox and can't modify it etc.
OpenCode's security problems are not limited to just permissioning agent actions. For a long time, simply starting OpenCode, and without requiring any other actions or prior configuration, allowed any website to run any code on your computer without any restrictions.
I dont want it doing damage to the sanbox either. I have half a dozen sandboxes running 24/7, there's definately a balance. My sandbox can't say dont run python for this type of task or similar conditional permissions. There are gaurdrails that the harness should handle well.
Also the harness should definately just have a solid permission structure for the average user that 1000% is not using a sandbox, most people I know naively do not and think these things are perfectly safe.
Yeah I came to the same conclusion, I migrated to run harnesses and my projects inside a container. However some kinds of development are really hard to run inside a container (like native mobile development or 3d game engines).
I have even been running my IDE in server/client mode with the server inside the container too (because of malicious extensions).
Even better, just limit what the agent can access. Don’t give it bash, give it the exact collection of tools you want it to be able to use. Limit it, it still works totally fine without full bash access. I gave my harness access to everything it needs through tool calls to specifically scoped actions and it works great
this is ultimately the right approach. You are going to get more mileage by giving your agent a specific set of tools and operations it can freely access and restrict the rest. Ive also found that any dynamic rejections pollute the context and may lead the agent to go find worse workarounds.
Put your agent in a restricted but fixed environment, and let it do its task. Anything else is mostly theatre and token burn.
Wrong. This is largely a solved problem with “auto review” permissions mode, which uses another LLM to review all tool calls that could potentially be dangerous. That solution isn’t perfect of course but it is reliable and usable at the point. This is available in Claude code, codex, copilot, and a bunch of others.
I find auto review mode mandatory and refuse to consider any harness that doesn’t support it.
Opencode doesn’t support it. And there are no good plugins to do it either, the most popular had only six stars last I checked.
> Their CEO spends all day on twitter talking about how much better he is than everyone. And how nobody can use agents like he can. Its quite obnoxious
I follow him on X, and I never got that impression. He has always come across as a skeptic IMO. Often wondering what others know that he doesn't due to practices that he hasn't found useful. I think yesterday he posted something suggesting that he doesn't follow usage patterns posted by others because the models have advanced to the point where any perceived improvements are by model rather than these patterns. I generally agree with this, and try not to get attached to any way of working with these things.
As for the usage of OpenCode, I have kept it to a minimum. V1 memory usage got a bit wild. Not sure if V2 has greatly improved. I have signed up for their Go plan a few times for times when I get tight on my other subs.
In the latest open code they changed the inference so now you getting more tokens per second it's around 300 to 400 per second code also they have provided no attention high for class with a near unlimited usage!
the guys are vibecoding and behaving like adhd company. on twitter creating 10x slop videos or new products when the opencode go dashboard was broken for days. opencode v2 itelf is broken just yesterday i said from main session to trigger a subagent of x type. but it overrode my provider of said model and used api key instead subscription causing me to spend a lot more then i wanted. i have a feeling everything is half assed. idk why i use it anymore, i guess the idea was to use one harnes because i use (to many) different models and providers but in the end i have native harrnes for all of them.
> Opencode V2 basically threw security out the window.
Is that about agent permissions/access? I actually prefer the approach where they ship an unlocked product and it’s on you to decide how to configure and deploy it. Claude Code and Codex ship complex, but partial sandboxes that could surprise a lot of users in what they allow.
This, I generally want gates for stuff that impacts outside world like terraform apply, gh pr merge etc but agents should have pretty free reign at sandbox level. And IMO if you only trust harness safety features you are doing something very wrong.
I use oh-my-pi, but I also need to use it to control a real browser I’ve warmed up so various sites think it’s a real person using it. There’s a very valid use case for this.
I am referring to "craftsman" software: it covers broad range of users flows well (from generic on-boarding, to very specific feature that rarely used), with generic user behavior modeling so it "feels" natural when reveal features progressively, with proper animation in-between to visually guide user's mental model of the software itself.
These are gone, not because the generic "enshittification", but because with the new-wave of software, there is no need to "model user behavior" or guide "user metal model" any more: user tells exactly what software they want, based on what their mental model is. There is no need for such craftsman software any more.
You are using that word wrong. What you described is just good UX. Craftsmanship is more about how the software looks like from the inside. It’s architecture, separation of concerns, test ability and malleability.
agreed! never understood the software world's ongoing obsession with twittersphere and how important some ppl take it, esp. when its full of so many dumbass opinions competing for space
> Also just use CDP, way too many browser solutions for agents when chrome ships with the ability to crontrol the browser at a low level.
> If you want to wrap it in a jsonrpc api for mcp, this is like 4 hrs or a work for an llm, point them at a pkg thats uses CDP like go-rod.
As someone who has vibe-coded my own such wrapper, I think that there is a lot of value in having a common interface for an agent to interact with CDP beyond just slinging together ad-hoc scripts until something works. CDP exposes a lot of features, but exposing them to an agent in a sane, consistent way takes a good bit of work.
It's also taken me a good bit longer than 4 hours by now. The features really add up over time.
Yeah but if you ask the agent, they are probably tripping over their ad-hoc scripts and working around little issues all the time. All that stuff adds up. Maybe my wrapper is a YAGNI-esque optimization on my part to avoid this, but I find it's also fun to work on these sorts of tools.
> Yeah but if you ask the agent, they are probably tripping over their ad-hoc scripts and working around little issues all the time
I have done this periodically - the last time, coincidentally, was yesterday because I'd upgraded node and forgotten to reinstall playwright-cli so I noticed agents jumping through hoops to use it.
So I started a review session, asked why that was happening, and then asked the agent to do a few tasks of its choice using it and report any issues. Besides installing playwright properly it suggested a few very minor procedural updates but that everything else was fine. This is unusual for agents because they'll nearly always find things to do. If they say something is fine it's usually a sign that's it's genuinely pretty good.
They have raised close to 1B and it seems financially they'd have raked in already a lot. Their subscriptions are gong quite well it seems – it's really cheap. So I am not sure they have lost the plot as an org for themselves. As for losing the plot as an org for its users – don't think that was the goal ever. Go to their GitHub to see for yourself. These guys inserted themselves into the "og" OpenCode and then left with the name because they had registered the domain name and made a much inferior TS clone and made it famous which is all it counts in today's time. Job done.
Pulumi and creating your own abstractions for it I think. That's most of what SST always looked like to me.
Unless you mean SST Console specifically, which I think is what GP is talking about. I didn't actually know that existed until just now but it actually looks pretty good. Don't think I've seen anything else quite like it before.
Shame they abandoned it. I like Pulumi honestly so anything that builds on top of it/improves its developer experience is great to see.
One example: I wanted to lock down code execution to run in bwrap. Claude made a nice extension for this, which correctly sandboxes shell commands. Then I asked it to break out, it said sure and dumped a config file in a subfolder and moved the whole session into that folder. Opencode picked up the new permissive config and “hello from inside opencode.txt” appeared on my desktop, all without any questions asked. It’s possible to lock down all these side doors, but opencode doesn’t even provide labels in the permission prompts for most of them. Still a nice package though, just not concerned with security
I agree to sandbox but opencode doesnt instruct you to use bwrap and many people dont even know what a sandbox is and you have the head of Opencode, dax, screaming about how dumb you are for caring about permissions while marketing to as many normies as possible.
These guys have been building amazing and useful products for years.
SST was a tool ahead of its time when it launched (I used it). It’s on version 3 and is now open source and community maintained. I’ve found it useful.
The OpenCode agent is very cool and they even host some of their own models via a subscription. I found it useful too.
I haven’t used this new one yet.
Have you built anything useful, or are you just a critic?
ok, last i used it, which was within the last 12 months, it was v3. i was wrong.
> is not maintained
they're running off investor money (maybe profitable now, i don't know), do you expect them to spend their runway to maintain a project that clearly isn't going to work as a business model?
they spent years, day and night (i'm sure) working as hard as possible to build the SST tool, which by the way, could be used for free. (they even built seed.run - a way to deploy your SST projects)
how many years of their work are we entitled to? i am going to keep defending these guys against critics. it is so easy to be a critic and so hard to build something from nothing.
people pay for the SST console. I would say that people are entitled to their work as part of the general expectation of how paying for services works.
vulnerabilities in SST have been reported to them. in April they announced [0] their intent to maintain and support SST. take a look at the commit activity graph [1]. I'm not saying that they are beholden to provide support for years for free. I'm saying that they are currently paid to support SST and are not providing it, and that for those who aren't paying they owe an acknowledgement.
You sound biased for someone who's hardly used the product.
Yes I put in 12 hrs a day, and usually running overnight as well in maybe half a dozen containers. I have billions of tokens going through Opencode a day as my main harness, for atleast a year.
I cannot accept their security model in V2 nor am I willing accept the nightmare and dangerous process of an upgrade. Where they basically tell the user, it ships with no permission model your plugins wont work, and we're making arbitrary changes all the time because we dont like how something is spelt/is plural that break your config. Typical vibe coded changes that are made for no rhyme or reason that are pushed without care so the user deal with the headache. V2 is literaly slop because they got obsessed with full on vibe coding.
Instead of having a version bump to V2 that cared about backwards compatibility they said screw the user, we have llms now just tell it to upgrade your config and ask it to make no mistakes, we provided a skill, good luck (oh, and by the way we no longer ship an easy way to confine the agent into a directory, thats on you good luck.. Ask the llm who now has access to your entire computer, make no mistakes).
This wouldnt be that big of a deal if I didnt have half a dozen different sets of configs with granular permissions, each with their own plugins for different clients Im working for. Its probably a multi day task to upgrade for any power user which shouldn't be the case for any piece of software's version bump. If Im going to take the time to upgrade it would be to a different harness all together.
Oh and by the way, when you turn off acess to the internet to a container and use a local model, session titles suddenly stop working.. Turn it back on and they start working even with no Opencode acct authenticated, because they're summarizing your conversations without your consent to derive session titles.... So who knows what they're swallowing up of your conversation data idk, I should put it behind an burp and see.
They have 2k issues on github but are shipping native UI frameworks nobody is going to use, for twitter likes.
They got a bunch of their users keys stolen and never took responsibility but rather blamed it on "distillers" instead of their own internal policies that lead to it. You can thank Dax for how every AI company scapegoats all their failures to distillers now whenever they fumble your credentials.
Thats too bad, its the equivalent of being a gun manufacturer and opting not to put safetys on the weapons you make. I dont really trust their team tbh. I'll stay on v1 for now until I find time to migrate to a different harness.
This is pretty alarming. I try to set up my containers to only have web access via specific provider MCPs and not just free reign to the open web. I would not want to use a harness trying to fight against me on what is already a very hard problem.
For me, I don't worry about what permissions I give my harness because it runs in a container where it can't harm anything unless the agent is being actively malicious and looking for sandbox exploits. I probably have bigger problems in that case.
I feel very validated to hear someone else rant about the v2 upgrade as a user of OpenCode. As a user it left me wondering
"If AI is so great at coding, and you're so good at using AI, why can't you make the changes you want in the v1 codebase gradually? Why can't you use AI to solve (not just mass close, actually FIX) the thousands of open issues you have on your GitHub?"
I think this is making me realize that I don't want to keep using OpenCode after using it as my daily driver for over a year (thousands of sessions).
And before anyone says "use Pi, it's great", I tried it and while it is fine, I think I fundamentally don't want to use a node-based tool ever again. Don't know why, they just always feel "slick" in all the least useful, most flashy ways. Anyone have any recommendations for the "sublime text 3"/nvim of stable and reliable harnesses?
> And before anyone says "use Pi, it's great", I tried it and while it is fine, I think I fundamentally don't want to use a node-based tool ever again. Don't know why, they just always feel "slick" in all the least useful, most flashy ways. Anyone have any recommendations for the "sublime text 3"/nvim of stable and reliable harnesses?
I really don't get it, since Pi has always been "vim of harness space" for me...
If you're keen to try something outside the terminal, give my harness dreamcoder.ai a whirl.
It runs on all 3 platforms on a machine of your choice, from your browser. It has a tiling window manager and it is a frontier coding harness. It is not written in node.
DreamCoder's built to be stable and run day in and day out. It's not perfect, because harnesses are exceptionally complex, but it works really well.
It is a _very_ different paradigm to what you have used in Pi/OpenCode TUI.
That it's on the same, level playing field as your favourite harness.
That's not so easy to achieve as you might think. Like all good software you can get 80% of the way to something quite quickly; it's the remaining 20% that takes a long time.
As for code generation, I'm not sure what you mean precisely?
For the way I work the #1 feature of a harness is quality code generation. It's nice that you can organize agents and whatnot, but if they all spit out suboptimal code for the task at hand that's not super relevant.
Ah, I think I misunderstood you. Sorry about that.
You can use any model you like with dreamcoder. The harness steers the models well, and they'll turn out the same quality code as the other frontier harnesses.
I use several: Qwen 3.8 27B for local, which is a cracked little model; a Codex sub (Astra, Sol); and the free monthly credits I get with my claude sub for UI/visual stuff.
yeah electron but small and fast. imo it’s nicer to have a desktop app than require a backend and I’m pretty sure with virtually zero changes you could ship it! especially if you’re already using rust.
> I fundamentally don't want to use a node-based tool ever again
These are not (nostly) written in JS/TS, so I think these are not node based.
https://github.com/openai/codex (this is rust but I don't absurd dependency depths rust seems to be getting, so in that aspect rust tools is becoming like npm tools)
As for how reliable these are, you'd have to try them and see for your own usage. I am sure there are more. Lighter ones, maybe less famous, but I have tried these and I am back to using Claude Code most of the time :D. Though I am trying to get familiar with Goose. However given a choice between OpenCode and others, I won't even bother with OpenCode – it's really that bad. Though I've their Go plan and that has been very userful for me (super vfm).
I also don't get the point of Pi. Pi either has to become Claude/Codex for most coding tasks and then what's the point it, otherwise a tiny harness (or script) replaces it, so I've also stopped using Pi. Apparently there Pi ports in other languages.
> I also don't get the point of Pi. Pi either has to become Claude/Codex for most coding tasks and then what's the point it, otherwise a tiny harness (or script) replaces it
Do you have an example of tiny harness?
I use Pi, without any plugins and it is enough for me, I don't need more. The only issue is that I don't like JS/TS (its place is in the web, not in terminal) and I would prefer something written in a compiled language (C or Rust preferably) or at least Java/Python.
Oh, and I would need integration with Github Copilot and Claude APIs (api pricing) so I could use what I get at work.
Comment below mentions hax, it looks promising (I don't need mcp, having "code mode" would be nice, but I'm still not sure if it is useful, and I would prefer it to not use JS).
I found this to be too minimal, but I would try again. Are there any power users of it that have written up? The thing is I do like plugin-esque paradigms.and I do need mcp functionality.
I used OpenCode a lot before, but I had to drop it and switch to Claude code (again). It’s missing a lot of features, especially session-to-session communication.
He runs Opencode and major stakeholder in the parent company. They don't too publish an org chart. Just because you dont give yourself the the Chief Executive title doesn't mean you aren't the Chief executive. Thr not a ceo thing is kind of a twitter meme too, hes the guy in charge.
> doesn't mean you aren't the Chief executive. Thr not a ceo thing is kind of a twitter meme
Jay V (@jayair) seems to be explicitly publicly listed as the CEO. Is that enough to make someone else who says they aren’t the CEO, not the CEO? I just don’t understand the continued insistence to use that specific terminology when the company says it isn’t how they’re organized?
I 've been using OpenCode Desktop before I migrated to OpenClaw. Haven't used v2 as much because of this. There were papercuts, but not as many as OpenClaw :p Still was happy with it and if not for OpenClaw being that good, I 'd likely still be using it.
> Their CEO spends all day on twitter talking about how much better he is than everyone.
Not that agreement proves anything but I am glad to know I was not the only one. Him, Boris, Theo (t3.gg or something) -- I look at their archetype and wonder if that is what you need for success and if indeed that is what you need, then perhaps I do not want success at all.
It's interesting because it's not strictly speaking needed.
You have people who have high visibility on social media without playing games or constantly exaggerating their accomplishments or work.
But you're right that people who do play these games and adopt your so called archetype end up with this much visibility.
I think the reality is as simple as grandiose claims or presentation or just being clickbait personified works to gain an audience. It's also why these people have a bigger ratio of fans to people who can't stand them.
The much rarer people who reach a large audience while remaining genuine tend to be much more universally liked. The ratio is better for sure at least.
it's been this way for years tbh. they made the whole web scene insufferable. the few gems have remained in the web scene, while everyone else moved to ai.
They see success publicly and seem to be drawn to it like flies on shit. I mean look at the drivel and pure propaganda that Garry Tan posts on X still. It's like money gives you a pass to be a raging lunatic.
>If you want to wrap it in a jsonrpc api for mcp, this is like 4 hrs or a work for an llm
isn't vibecoding a throwaway integration like this what you're criticizing opencode for here? why would i waste my tokens on that when opencode will give it to me already built?
cdp is maintained by Google.. You realize that.. Right?
A jsonRPC wrapper isnt exactly the critical part here. Also not advocating for that, just use cdp lol. There are also well maintained wrappers of cdp that existed long before Ai/AnomalSlop that are maintained by incredible devs, like the one I mentioned go-rod maintained by yasmood.. Whos actually an expert on browsers/cdp unlike anaomoly who has zero background on it. .
Adding a jsonRPC wrapper to existing and maintained solutions for tool annotations isnt exactly that big of a deal... And you can then choose your security model rather than allow dax to choose for you, who doesnt give a fk about security.
1. Rune
2. Hex
3. Browse
4. Nightmail
5. Deadbolt
6. Strobe
7. Phonecall
8. Justice
9. V2 (Not 3 months old yet)
10. Flicks
11. Hokum
12. Fiend
13. Isocast
14. Clip
15. Grim
16. Bellhop
17. bank
18. Blackout
19. Text.js
20. Jobmail
21. Bane (If want to exclude V2)
---
Operation CheepSeek (The latest phase at least)
Operation GLM
OpenTunnel (Just out the date range)
There are more repos, and some of these are internal use, but are published under their umbrella
(Nothing against Anomaly, I just follow them on twitter and see anouncements for new stuff regularly enough to think that it was possible)
Yeah I've been somewhat disappointed with opencode lately (not as disappointed as I am in codex though)... the issues and PRs on GitHub are continuously accumulating and they don't seem to have any structured/efficient way to deal with it. Obvious issues with obvious fixes, and zero disagreement are sitting so long that they just get auto closed after 1 month.
Codex has same problem that oh-my-pi has that gui feels too much magical box where you put prompts and software comes out. Opencode nailed balance between output and traces pretty well but sadly they also are going to that magic box direction with latest versions.
I keep asking my own agents if prompt injection is a solved problem. By and large the answer seems to be ‘no, anything a model sees can be used to steer it and enters the context’.
I know you can forward raw output through subagents to summarize content and mitigate the risk - but most projects don’t do that.
So why would anyone risk browser or email to agents where your secrets are at their most vulnerable. Serious question, looking for someone who knows this stuff and can reassure me (or not) on the real risks.
A friend who works in infosec was absolutely floored when I asked him how he was sandboxing the agents he runs on his main laptop. I guesss I'm finished being surprised but god damn.
The functionality seems to overlap with the Playwright MCP server [0]. I've been using this instead of the built in browsers in Claude Code and Codex since it lets you grant access to pages and sessions that are already open. Of course one should be weary of the lethal trifecta when using these tools and use separate profiles for them.
It's deferent; with a plugin in firefox, edge, chrome, chromium; I can bring an agent into it. Secure contexts like outlook I can have the agent scroll through my messages.
MCP servers to me seems like a dead man walking technology and I don't understand the point of it.
It adds security but...
- adds a single service which is another attack surface that can be compromised.
- adds another system that requires maintenance and securing to prevent compromise.
- takes machine resources
- duplicates a subset of what can easily be harness functionality
- duplicates services and security (usually) of existing APIs, or adds another layer on top.
Whereas ultimately an AI harness
- will improve and optimise its own skills performing the same MCP functions
- can use plugins
- can have harness centralised (common) security controls and behaviours
MCP becomes a false economy for our current AI implementations, it's real purpose seems to be a crutch for small models that struggle to write their own skills that use existing APIs. The thing is this is a problem that is rapidly disappearing, but could also be performed by secured harness plugins.
I am going to push back on this a bit as someone who also has scepticism (bloated spec, concepts deprecated already, etc.) but ultimately think it solves a few gnarly problems whilst admittedly -- it's early days -- creating a few more.
The first argument in favour of MCP is that large enterprises like it. It's a controlled surface to give access to things that may never have had an API before. Yeah, OK, that feels like a "loser takes all" moment but it does at least give you access to something you did not have before.
It also resolves a problem: monolithic apps without a CLI/server infrastructure can now expose a simple MCP endpoint and immediately get it to work locally with their own agents. That has value too.
The second is that it tries to solve thorny things like letting the server ask for information by way of a form-like construct. That is actually quite handy for servers that do asynchronous or slow work, and that may require feedback from you over time. Doing this with start-stopping a tool harness can also work, but it's fiddly because a context compaction event can happen which might lose state; not all harnesses handle this well; LLMs may get confused; etc.
Finally you can inject, in a good harness anyway, some or all of these tools into your system so they become part of the fabric. You can build your own little cut-down surface area so by removing shell access you let your AI call Foo and Bar MCP calls and nothing else. That also is useful.
But yes: you're right fundamentally. It's SOAP/RPC all over again, and CLI+skill should be the default for your own tooling if you can get away with it.
- they necessarily work through the context window and therefore don’t compose;
Meaning, if your agent has a skill that needs to run five MCP tools in a certain way, it can’t just script that away the first time it sees this, and have that flow be fast and token efficient from then on.
MCPs are a 2025 tech meant to solve 2025 AI problems.
And then your agent goes rogue through your harness, deletes your database, sends your passwords to an e-mail and wipes your filesystem because it saw a prompt injected somewhere on a random webpage you asked it to search for. Then you will wish for a deterministic MCP server you can fully lock down and control what the agent can do with your infrastructure.
I have been using Open Code with Big Pickle, basically I have downloaded it to my desktop and used it with a local host. So far it has cost me nothing. It is a way to do things for very little money. I have run out of time a couple time, but I just go for the next. I don't even look at X or Twitter.
Opencode has a lot of issues and I've tried all the alternatives but ultimately it's the one I go back to. You have to treat it like untrusted code full of telemetry and data leaks like the infamous title generations. Ask your agent to craft a config disabling everything, and then run it in a sandbox without direct network access with a proxy for whitelisted domains. But that applies to any agent really. The tabs in the v2.0 TUI are a game changer for me so it looks like I'll keep using it for now.
So with that said, how do you feel about an extension that explicitly enables Opencode to execute tasks on your behalf, as if it was you, using your credentials and 2FA-authenticated sessions?
I have been using opencode for quite along time even before they had an official ios port and think it's the most flexible harness for the best bang for your buck. 10 quid with practicality unlimited usage for side projects and personal agent stuff.
This is promising. My primary use of Codex is browser automation, and I perpetually keep running out of tokens for menial tasks. OpenCode + Spark Muse will make my life so much easier.
It is a chromium extension and Firefox is not a chromium-based browser.
Like or dislike Firefox, omitting it from a list called "real browser" just looks amateur or sloppy. How about "real chromium browser"? Was it really that hard?
I added both browser and kitesurf today, so I'm still very early assessing which I'll stick with. But nice to see two very lightweight browser systems show up.
I was just looking into these "chat with ai agent about a webpage" browser extensions/projects/forks ... can someone help me sort out this field? it seems really busy.
What I really want is like something that can search, use my own openai compatible server with, and I think that's it? I dunno what else these things do.
Looks cool.
I slopped something similar but with also full desktop control in addition to the browser debugger extension.
Because I was not satisfied with openai computer use being limited and refusing to interact with menu bar/task tray.
https://machinecontrol.dev/
Use it daily with Claude and codex.
I found agent browsing to be an issue when I started using them months ago. Built a Firefox extension back then that uses the browser instance that I use, the way I use it, so there's really no difference to a site.
I don't understand the need for this. A strong reasoning model already knows how to use Powershell to launch my browser with my profile. gpt-6.1-terra will build a whole goddamn .NET project with playwright just to pull the showtimes from the local cineplex.
Shell execute and apply patch are the only two tools you need if your reasoning model is powerful enough.
What I'm trying to say is that the frontier models will already do this. At least OpenAI's models will. It launches chromium edge headlessly using my userland profile. From the browsers perspective it's me using it.
Funny, I just vibecoded a Chrome extension that does exactly that for my agent harness. I wonder if at OpenAI they too just told GPT “Let’s create a browser and Chrome extension for browser use, when done publish it on our website and post it to HN”. Because the agent itself is the user it worked until it’s happy fulfilling my request, modifying and adding functions when itself needs.
I work in a large engineering org, and I've never heard anyone using these for work, while opencode is pretty much the default open-source agent harness.
been using opencode since it came out in like april 2025 (i could be wrong but i feel like this is when the whole charm.sh stuff happened with them) and never felt like i ever had a reason to use anything else. i've tried claude code and pi and stuff but all of them just felt quite unpolished compared to the opencode TUI.
i spend a lot of my time just copying ideas from their server runtime into https://wingman.actor.
OpenCode comes with the cost too. The cost of the hosted version is on par with cloud hosted proprietary models. And we cannot simply run the LLMs on every laptop. Given this, how do we us the browse efficiently?
Yea it was a pity when opencode limited the token usage and told us to subscribe to their opencode go service. Like openai, it's clearly angling for commercial viability status under the guise of being 'open'.
You can implement this as a skill in any AI agent in about 5 minutes, just ask it to make this and it will. I've done this with Chrome and Firefox five different ways, not counting MCP and pre-existing extensions. I basically have one skill that calls four different browser skills depending on how hardcore the bot detection is. New browser profile each session. 1Password and a custom app let me prompt for auth on any machine.
Tangent: Xvfb makes it crazy easy to implement headless computer use for any app. Add SSH or VNC for remote access. Open weights easily use the x11 tools to control the GUIs, makes testing and automation a breeze.
It's all right when it works, but it has a habit of disconnecting; you have to rerun 'browse pair'. This MCP should hopefully become unnecessary once OpenCode's integrated browser tool gains persistent sessions. https://github.com/anomalyco/opencode/pull/63002
Opencode vibe codes and releases like 3 projects a week at this points. Cool, just know its throwaway software to them. They seem really reckless when it comes to security. Opencode V2 basically threw security out the window. Theyve completely lost the plot as an org.
Their CEO spends all day on twitter talking about how much better he is than everyone. And how nobody can use agents like he can. Its quite obnoxious
Also just use CDP, way too many browser solutions for agents when chrome ships with the ability to control the browser at a low level.
If you want to wrap it in a jsonrpc api for mcp, this is like 4 hrs or a work for an llm, point them at a pkg thats uses CDP like go-rod.
They're totally right to not bother with security inside a harness, its just a terrible approach that creates permission fatigue and bad UX. Claude asking me to approve a 50 line arcane bash command doesn't help at all. Pi also has the same philosophy of not enforcing security within the harness.
Security should be enforced outside the harness with a sandbox. You can define what's allowed and what's not in one place and have strong enforcement guarantees instead of ad-hoc permission requests every minute.
> Security should be enforced outside the harness with a sandbox. You can define what's allowed and what's not in one place and have strong enforcement guarantees instead of ad-hoc permission requests every minute.
I agree in principle, but in practice most sandboxing technology (like bwrap) is too static for this use case: You grant access once during startup and can't extend/revoke without a restart. That's inconvenient when you realize halfway through your session that you need additional permissions, and it's insufficient for permissions that are only briefly needed (for the duration of one tool call).
A harness that controls the sandbox (e.g. for each individual tool invocation) can manage this much more effectively.
I almost never have to extend permissions of my sandbox. By default it can read/write the project tree, and read most files on the machine with specific secret locations excluded. If I do have to expand permissions and restart, I just resume the session.
A harness that controls the sandbox is exactly what I’m after. Tools don’t always need to share the same space, and I can isolate work much better if separate toolchains are isolated in sandboxes. My biggest gripe with the major harnesses is how they sandbox tool calls.
100%.
I have my agent request permissions for things like a Cloudflare API key (or I guide it or just come up with the request myself). These go in a simple queue, with justifications.
I review them and then have another agent with far more permissions create the tokens, with expiry and so on, and then hand the API tokens back.
The original agent doesn’t even run as my main user; it sits on a computer in a restricted account (easy to do with either macOS or Linux, or even Windows these days).
It has taken some rather long prompts to get it to self-report on itself when it tries to workaround.
Security when calling external services is much more than just FS permissions, sandbox doesn’t protect at all what you push to github or call API with
Doesn't change his argument that sandboxing should be done outside the harness. You can also handle network access etc externally.
You really wanting handle this at the sanbox layer? I have 100 different containers configurations, only a few differnet harness configurations.
You're just being stubborn, the best UX is through the harness. Most users dont even know how to configure a sandbox properly and dont even use one.
Ive known quite a few naive people that have gotten their home directories wiped and really gotten burned. Many users of these products dont even know how to use a terminal and these harness companies are marketing to the average joe.
> Most users dont even know how to configure a sandbox properly and dont even use one.
This is their problem. But running and screaming "ai deleted my database" is all on them, not on LLM.
I have a sandbox script and a simple wrapper that runs any command in that sandbox thus that "pi" -> without sandbox, "pisb" -> pi in the sandbox. I practically only use pisb.
This separation doesn't mean it must result in bad UX. It just needs somebody to make a nice UI/TUI to configure the sandbox and when running that it does its setup and open whatever command you want.
The problem with the harness doing sandboxing is that from my experience the simply break all the time leading to users ignoring properly configuring it and using abominations like codex' `Approve for me` where IT ASKS A SEPARATE AI and not the user.
My main issue is these are being marketed to average joes. You wont see anything on these harnesses documentation, opencode et al, who say it should be used in a sandbox. The people making them, opencode et al, scream everyday about how you're holding yourself back by using permissions without a single mention of sandbox.
Hundreds of thousands of people are using them naively, if they shipped with common sense strictish permissions instead of none at all, like what opencode opted to do it would be safer for everyone. Its much easier for a normal non programmer to configure a simple config file than it is for them to learn how to properly setup a container/vm, they will just not do the latter 9/10.
Honestly, a harness should always install heavily gaurded, remember when Grok code/build was instantly indexing your home directory and sending it back to HQ on the first request it made..
Like I wrote above I generally agree that the sandboxing should by enabled "by default" and should not be an extra opt-in that you have to know about first.
That being said I still believe this should be separated and composable at an application level and power users should be able to switch out the sandbox. This way you don't have to trust the harness since it doesn't know anything about the sandbox and can't modify it etc.
You can make restricted GitHub accounts and API tokens.
OpenCode's security problems are not limited to just permissioning agent actions. For a long time, simply starting OpenCode, and without requiring any other actions or prior configuration, allowed any website to run any code on your computer without any restrictions.
I dont want it doing damage to the sanbox either. I have half a dozen sandboxes running 24/7, there's definately a balance. My sandbox can't say dont run python for this type of task or similar conditional permissions. There are gaurdrails that the harness should handle well.
Also the harness should definately just have a solid permission structure for the average user that 1000% is not using a sandbox, most people I know naively do not and think these things are perfectly safe.
Yeah I came to the same conclusion, I migrated to run harnesses and my projects inside a container. However some kinds of development are really hard to run inside a container (like native mobile development or 3d game engines).
I have even been running my IDE in server/client mode with the server inside the container too (because of malicious extensions).
Even better, just limit what the agent can access. Don’t give it bash, give it the exact collection of tools you want it to be able to use. Limit it, it still works totally fine without full bash access. I gave my harness access to everything it needs through tool calls to specifically scoped actions and it works great
this is ultimately the right approach. You are going to get more mileage by giving your agent a specific set of tools and operations it can freely access and restrict the rest. Ive also found that any dynamic rejections pollute the context and may lead the agent to go find worse workarounds.
Put your agent in a restricted but fixed environment, and let it do its task. Anything else is mostly theatre and token burn.
Wrong. This is largely a solved problem with “auto review” permissions mode, which uses another LLM to review all tool calls that could potentially be dangerous. That solution isn’t perfect of course but it is reliable and usable at the point. This is available in Claude code, codex, copilot, and a bunch of others.
I find auto review mode mandatory and refuse to consider any harness that doesn’t support it.
Opencode doesn’t support it. And there are no good plugins to do it either, the most popular had only six stars last I checked.
> Their CEO spends all day on twitter talking about how much better he is than everyone. And how nobody can use agents like he can. Its quite obnoxious
I follow him on X, and I never got that impression. He has always come across as a skeptic IMO. Often wondering what others know that he doesn't due to practices that he hasn't found useful. I think yesterday he posted something suggesting that he doesn't follow usage patterns posted by others because the models have advanced to the point where any perceived improvements are by model rather than these patterns. I generally agree with this, and try not to get attached to any way of working with these things.
As for the usage of OpenCode, I have kept it to a minimum. V1 memory usage got a bit wild. Not sure if V2 has greatly improved. I have signed up for their Go plan a few times for times when I get tight on my other subs.
He is likely talking about Dax who trolls a lot but is not the CEO.
Yes, Dax. I was under an impression he is the CEO.
Dax is the CEO of HTMX
In the latest open code they changed the inference so now you getting more tokens per second it's around 300 to 400 per second code also they have provided no attention high for class with a near unlimited usage!
the guys are vibecoding and behaving like adhd company. on twitter creating 10x slop videos or new products when the opencode go dashboard was broken for days. opencode v2 itelf is broken just yesterday i said from main session to trigger a subagent of x type. but it overrode my provider of said model and used api key instead subscription causing me to spend a lot more then i wanted. i have a feeling everything is half assed. idk why i use it anymore, i guess the idea was to use one harnes because i use (to many) different models and providers but in the end i have native harrnes for all of them.
> Opencode V2 basically threw security out the window.
Is that about agent permissions/access? I actually prefer the approach where they ship an unlocked product and it’s on you to decide how to configure and deploy it. Claude Code and Codex ship complex, but partial sandboxes that could surprise a lot of users in what they allow.
This, I generally want gates for stuff that impacts outside world like terraform apply, gh pr merge etc but agents should have pretty free reign at sandbox level. And IMO if you only trust harness safety features you are doing something very wrong.
I use oh-my-pi, but I also need to use it to control a real browser I’ve warmed up so various sites think it’s a real person using it. There’s a very valid use case for this.
Their marketing is very effective. And marketing is the only thing matters in software industry going forward. Making good software is not.
Is there a harness that is good software?
Every single one I looked into looks like low quality vibecoded typescript.
https://github.com/charmbracelet/crush
Worth mentioning this is one of its developers.
I had good experiences with deepseek-harness and open-mercato/cezar.
What is good software. Something that works / solves a problem? Or something else?
I am referring to "craftsman" software: it covers broad range of users flows well (from generic on-boarding, to very specific feature that rarely used), with generic user behavior modeling so it "feels" natural when reveal features progressively, with proper animation in-between to visually guide user's mental model of the software itself.
These are gone, not because the generic "enshittification", but because with the new-wave of software, there is no need to "model user behavior" or guide "user metal model" any more: user tells exactly what software they want, based on what their mental model is. There is no need for such craftsman software any more.
You are using that word wrong. What you described is just good UX. Craftsmanship is more about how the software looks like from the inside. It’s architecture, separation of concerns, test ability and malleability.
Interesting. Where do you see their marketing that impresses you?
I mainly read Hacker News. So not surprised I've missed it.
But, I'm curious because I love to study good marketing.
agreed! never understood the software world's ongoing obsession with twittersphere and how important some ppl take it, esp. when its full of so many dumbass opinions competing for space
Odd thing to say after the MSFT and FAA outages but sadly you’re probably right. The prophecy of Idiocracy is coming to pass.
> Also just use CDP, way too many browser solutions for agents when chrome ships with the ability to crontrol the browser at a low level.
> If you want to wrap it in a jsonrpc api for mcp, this is like 4 hrs or a work for an llm, point them at a pkg thats uses CDP like go-rod.
As someone who has vibe-coded my own such wrapper, I think that there is a lot of value in having a common interface for an agent to interact with CDP beyond just slinging together ad-hoc scripts until something works. CDP exposes a lot of features, but exposing them to an agent in a sane, consistent way takes a good bit of work.
It's also taken me a good bit longer than 4 hours by now. The features really add up over time.
I use playwright-cli any a skill that's about 150 lines. Seems to have been working well for about 6 months now.
Yeah but if you ask the agent, they are probably tripping over their ad-hoc scripts and working around little issues all the time. All that stuff adds up. Maybe my wrapper is a YAGNI-esque optimization on my part to avoid this, but I find it's also fun to work on these sorts of tools.
> Yeah but if you ask the agent, they are probably tripping over their ad-hoc scripts and working around little issues all the time
I have done this periodically - the last time, coincidentally, was yesterday because I'd upgraded node and forgotten to reinstall playwright-cli so I noticed agents jumping through hoops to use it.
So I started a review session, asked why that was happening, and then asked the agent to do a few tasks of its choice using it and report any issues. Besides installing playwright properly it suggested a few very minor procedural updates but that everything else was fine. This is unusual for agents because they'll nearly always find things to do. If they say something is fine it's usually a sign that's it's genuinely pretty good.
I used gpt 6.1 sol / medium for the review.
Took me about half a day.
Granted, Ive neen building browser automation tools on top of cdp for many years now.
They have raised close to 1B and it seems financially they'd have raked in already a lot. Their subscriptions are gong quite well it seems – it's really cheap. So I am not sure they have lost the plot as an org for themselves. As for losing the plot as an org for its users – don't think that was the goal ever. Go to their GitHub to see for yourself. These guys inserted themselves into the "og" OpenCode and then left with the name because they had registered the domain name and made a much inferior TS clone and made it famous which is all it counts in today's time. Job done.
They also seem to have abandoned their pre-OpenCode success SST with no acknowledgement to paying customers.
Love SST, what are the alternatives?
Pulumi and creating your own abstractions for it I think. That's most of what SST always looked like to me.
Unless you mean SST Console specifically, which I think is what GP is talking about. I didn't actually know that existed until just now but it actually looks pretty good. Don't think I've seen anything else quite like it before.
Shame they abandoned it. I like Pulumi honestly so anything that builds on top of it/improves its developer experience is great to see.
https://alchemy.dev and you won't want to use anything else. don't waste time with pulumi or terraform
alchemy.run*
Both opencode and alchemy.run are big proponents of effect.
ops, sorry
> Opencode V2 basically threw security out the window.
Please elaborate. What is the problem if I only use OpenCode v2 (without extensions like Browse).
One example: I wanted to lock down code execution to run in bwrap. Claude made a nice extension for this, which correctly sandboxes shell commands. Then I asked it to break out, it said sure and dumped a config file in a subfolder and moved the whole session into that folder. Opencode picked up the new permissive config and “hello from inside opencode.txt” appeared on my desktop, all without any questions asked. It’s possible to lock down all these side doors, but opencode doesn’t even provide labels in the permission prompts for most of them. Still a nice package though, just not concerned with security
> lock down code execution to run in bwrap. Claude made a nice extension
I think you're looking at it wrong. I run opencode itself inside bwrap (atop other layers) so anything it runs it's already sandboxed.
I agree to sandbox but opencode doesnt instruct you to use bwrap and many people dont even know what a sandbox is and you have the head of Opencode, dax, screaming about how dumb you are for caring about permissions while marketing to as many normies as possible.
How on earth can you open up your defence with “claude made a” ????
>Their CEO spends all day on twitter talking about how much better he is than everyone. And how nobody can use agents like he can. Its quite obnoxious
What are some specific examples of this?
These guys have been building amazing and useful products for years.
SST was a tool ahead of its time when it launched (I used it). It’s on version 3 and is now open source and community maintained. I’ve found it useful.
The OpenCode agent is very cool and they even host some of their own models via a subscription. I found it useful too.
I haven’t used this new one yet.
Have you built anything useful, or are you just a critic?
SST is on version 4, and is not maintained. There is the beginnings of a community fork due to not getting a response from the maintainers for months.
> SST is on version 4
ok, last i used it, which was within the last 12 months, it was v3. i was wrong.
> is not maintained
they're running off investor money (maybe profitable now, i don't know), do you expect them to spend their runway to maintain a project that clearly isn't going to work as a business model?
they spent years, day and night (i'm sure) working as hard as possible to build the SST tool, which by the way, could be used for free. (they even built seed.run - a way to deploy your SST projects)
how many years of their work are we entitled to? i am going to keep defending these guys against critics. it is so easy to be a critic and so hard to build something from nothing.
people pay for the SST console. I would say that people are entitled to their work as part of the general expectation of how paying for services works.
vulnerabilities in SST have been reported to them. in April they announced [0] their intent to maintain and support SST. take a look at the commit activity graph [1]. I'm not saying that they are beholden to provide support for years for free. I'm saying that they are currently paid to support SST and are not providing it, and that for those who aren't paying they owe an acknowledgement.
[0] https://x.com/vimtor/status/2039039247286501688 [1] https://github.com/anomalyco/sst/graphs/commit-activity
You sound biased for someone who's hardly used the product.
Yes I put in 12 hrs a day, and usually running overnight as well in maybe half a dozen containers. I have billions of tokens going through Opencode a day as my main harness, for atleast a year.
I cannot accept their security model in V2 nor am I willing accept the nightmare and dangerous process of an upgrade. Where they basically tell the user, it ships with no permission model your plugins wont work, and we're making arbitrary changes all the time because we dont like how something is spelt/is plural that break your config. Typical vibe coded changes that are made for no rhyme or reason that are pushed without care so the user deal with the headache. V2 is literaly slop because they got obsessed with full on vibe coding.
Instead of having a version bump to V2 that cared about backwards compatibility they said screw the user, we have llms now just tell it to upgrade your config and ask it to make no mistakes, we provided a skill, good luck (oh, and by the way we no longer ship an easy way to confine the agent into a directory, thats on you good luck.. Ask the llm who now has access to your entire computer, make no mistakes).
This wouldnt be that big of a deal if I didnt have half a dozen different sets of configs with granular permissions, each with their own plugins for different clients Im working for. Its probably a multi day task to upgrade for any power user which shouldn't be the case for any piece of software's version bump. If Im going to take the time to upgrade it would be to a different harness all together.
Oh and by the way, when you turn off acess to the internet to a container and use a local model, session titles suddenly stop working.. Turn it back on and they start working even with no Opencode acct authenticated, because they're summarizing your conversations without your consent to derive session titles.... So who knows what they're swallowing up of your conversation data idk, I should put it behind an burp and see.
They have 2k issues on github but are shipping native UI frameworks nobody is going to use, for twitter likes.
They got a bunch of their users keys stolen and never took responsibility but rather blamed it on "distillers" instead of their own internal policies that lead to it. You can thank Dax for how every AI company scapegoats all their failures to distillers now whenever they fumble your credentials.
OpenCode has never been concerned about security, and has explicitly said their goal is agent UX, not security.
This is from September 2025: https://github.com/anomalyco/opencode/issues/2632
Issue comment: "In my opinion very strange that Opencode allows any actions by default."
Comments from OpenCode devs:
"This was done very intentionally by the team they felt like this was a better UX and better default when they made the decision"
"most people initially think they want permissions and then actually prefer not having them so that's why our defaults work this way"
"I use it everyday in real projects at work, so do my coworkers, none of us use permissions"
Thats too bad, its the equivalent of being a gun manufacturer and opting not to put safetys on the weapons you make. I dont really trust their team tbh. I'll stay on v1 for now until I find time to migrate to a different harness.
Imo if you’re not running harnesses on sandbox you are doing something wrong and you absolutely should not trust harness safety features.
This is pretty alarming. I try to set up my containers to only have web access via specific provider MCPs and not just free reign to the open web. I would not want to use a harness trying to fight against me on what is already a very hard problem.
For me, I don't worry about what permissions I give my harness because it runs in a container where it can't harm anything unless the agent is being actively malicious and looking for sandbox exploits. I probably have bigger problems in that case.
sounds like a different product would be a better fit for you
I feel very validated to hear someone else rant about the v2 upgrade as a user of OpenCode. As a user it left me wondering
"If AI is so great at coding, and you're so good at using AI, why can't you make the changes you want in the v1 codebase gradually? Why can't you use AI to solve (not just mass close, actually FIX) the thousands of open issues you have on your GitHub?"
I think this is making me realize that I don't want to keep using OpenCode after using it as my daily driver for over a year (thousands of sessions).
And before anyone says "use Pi, it's great", I tried it and while it is fine, I think I fundamentally don't want to use a node-based tool ever again. Don't know why, they just always feel "slick" in all the least useful, most flashy ways. Anyone have any recommendations for the "sublime text 3"/nvim of stable and reliable harnesses?
> And before anyone says "use Pi, it's great", I tried it and while it is fine, I think I fundamentally don't want to use a node-based tool ever again. Don't know why, they just always feel "slick" in all the least useful, most flashy ways. Anyone have any recommendations for the "sublime text 3"/nvim of stable and reliable harnesses?
I really don't get it, since Pi has always been "vim of harness space" for me...
If you're keen to try something outside the terminal, give my harness dreamcoder.ai a whirl.
It runs on all 3 platforms on a machine of your choice, from your browser. It has a tiling window manager and it is a frontier coding harness. It is not written in node.
DreamCoder's built to be stable and run day in and day out. It's not perfect, because harnesses are exceptionally complex, but it works really well.
It is a _very_ different paradigm to what you have used in Pi/OpenCode TUI.
it is a frontier coding harness
What exactly does that mean, these days? Which harnesses are not frontier, and what does it do much better (on the code generation front) than others?
That it's on the same, level playing field as your favourite harness.
That's not so easy to achieve as you might think. Like all good software you can get 80% of the way to something quite quickly; it's the remaining 20% that takes a long time.
As for code generation, I'm not sure what you mean precisely?
For the way I work the #1 feature of a harness is quality code generation. It's nice that you can organize agents and whatnot, but if they all spit out suboptimal code for the task at hand that's not super relevant.
Ah, I think I misunderstood you. Sorry about that.
You can use any model you like with dreamcoder. The harness steers the models well, and they'll turn out the same quality code as the other frontier harnesses.
I use several: Qwen 3.8 27B for local, which is a cracked little model; a Codex sub (Astra, Sol); and the free monthly credits I get with my claude sub for UI/visual stuff.
why browser and not tauri?
I've not heard of/used Tauri, but I'm guessing it's some sort of Electron-shaped thing?
Browsers are everywhere, and you can point your phone at it. It also means the backend's a REST API, which has its own advantages.
yeah electron but small and fast. imo it’s nicer to have a desktop app than require a backend and I’m pretty sure with virtually zero changes you could ship it! especially if you’re already using rust.
I'll definitely take a look at Tauri. Thanks for turning me on to it.
> I fundamentally don't want to use a node-based tool ever again
These are not (nostly) written in JS/TS, so I think these are not node based.
https://github.com/openai/codex (this is rust but I don't absurd dependency depths rust seems to be getting, so in that aspect rust tools is becoming like npm tools)
https://github.com/nousresearch/hermes-agent
https://github.com/charmbracelet/crush
https://github.com/aaif-goose/goose https://aaif.io/projects
https://github.com/Aider-AI/aider kinda the original one
As for how reliable these are, you'd have to try them and see for your own usage. I am sure there are more. Lighter ones, maybe less famous, but I have tried these and I am back to using Claude Code most of the time :D. Though I am trying to get familiar with Goose. However given a choice between OpenCode and others, I won't even bother with OpenCode – it's really that bad. Though I've their Go plan and that has been very userful for me (super vfm).
People have recommended me these but I have not used any of these: https://github.com/SWE-agent/SWE-agent
I also don't get the point of Pi. Pi either has to become Claude/Codex for most coding tasks and then what's the point it, otherwise a tiny harness (or script) replaces it, so I've also stopped using Pi. Apparently there Pi ports in other languages.
> I also don't get the point of Pi. Pi either has to become Claude/Codex for most coding tasks and then what's the point it, otherwise a tiny harness (or script) replaces it
Do you have an example of tiny harness?
I use Pi, without any plugins and it is enough for me, I don't need more. The only issue is that I don't like JS/TS (its place is in the web, not in terminal) and I would prefer something written in a compiled language (C or Rust preferably) or at least Java/Python.
Oh, and I would need integration with Github Copilot and Claude APIs (api pricing) so I could use what I get at work.
Comment below mentions hax, it looks promising (I don't need mcp, having "code mode" would be nice, but I'm still not sure if it is useful, and I would prefer it to not use JS).
I really like maki [1]. It's been very well put together and I like its focus on speed and simplicity.
[1]: https://maki.sh/
Try hax (https://usehax.dev/). Minimal, not flashy, plain old school C.
I found this to be too minimal, but I would try again. Are there any power users of it that have written up? The thing is I do like plugin-esque paradigms.and I do need mcp functionality.
I used OpenCode a lot before, but I had to drop it and switch to Claude code (again). It’s missing a lot of features, especially session-to-session communication.
He's not the CEO (although everyone thinks he is)
He runs Opencode and major stakeholder in the parent company. They don't too publish an org chart. Just because you dont give yourself the the Chief Executive title doesn't mean you aren't the Chief executive. Thr not a ceo thing is kind of a twitter meme too, hes the guy in charge.
>>> Their CEO
> doesn't mean you aren't the Chief executive. Thr not a ceo thing is kind of a twitter meme
Jay V (@jayair) seems to be explicitly publicly listed as the CEO. Is that enough to make someone else who says they aren’t the CEO, not the CEO? I just don’t understand the continued insistence to use that specific terminology when the company says it isn’t how they’re organized?
I 've been using OpenCode Desktop before I migrated to OpenClaw. Haven't used v2 as much because of this. There were papercuts, but not as many as OpenClaw :p Still was happy with it and if not for OpenClaw being that good, I 'd likely still be using it.
I'm glad you made that comment because I didn't look very closely at the URL and thought this was an OpenAI product.
> Their CEO spends all day on twitter talking about how much better he is than everyone.
Not that agreement proves anything but I am glad to know I was not the only one. Him, Boris, Theo (t3.gg or something) -- I look at their archetype and wonder if that is what you need for success and if indeed that is what you need, then perhaps I do not want success at all.
It's interesting because it's not strictly speaking needed.
You have people who have high visibility on social media without playing games or constantly exaggerating their accomplishments or work.
But you're right that people who do play these games and adopt your so called archetype end up with this much visibility.
I think the reality is as simple as grandiose claims or presentation or just being clickbait personified works to gain an audience. It's also why these people have a bigger ratio of fans to people who can't stand them.
The much rarer people who reach a large audience while remaining genuine tend to be much more universally liked. The ratio is better for sure at least.
No, they just all know how to game the different socials and they do, even if it's obnoxious or cringe.
> if indeed that is what you need, then perhaps I do not want success at all.
what would success bring you?
if what you want is success at all costs however, that is where you have to start
it's been this way for years tbh. they made the whole web scene insufferable. the few gems have remained in the web scene, while everyone else moved to ai.
They see success publicly and seem to be drawn to it like flies on shit. I mean look at the drivel and pure propaganda that Garry Tan posts on X still. It's like money gives you a pass to be a raging lunatic.
Highly recommend "The Nerd Reich" [0].
[0] https://www.thenerdreich.com/
> Opencode V2 basically threw security out the window
Yes. I was shocked to suddenly discover everything was "auto permit all"
Then i discovered that all my pinned sessions were lost
Then you read on X that they vibing some sort of UI library or whatever the flavour of the week is
It a shame because they getting nice deals on tokens. You can churn billions of tokens per month
Just grow up already...
Unfortunately it takes years to grow up
>Also just use CDP
isn't that what this is?
>If you want to wrap it in a jsonrpc api for mcp, this is like 4 hrs or a work for an llm
isn't vibecoding a throwaway integration like this what you're criticizing opencode for here? why would i waste my tokens on that when opencode will give it to me already built?
I purged OpenCode (and KiloCode) after I realized the security situation (on code and mentality level)
>vibe codes and releases like 3 projects a week
https://opencode.ai/product
What?
>just know its throwaway software to them
>this is like 4 hrs or a work for an llm
Ah, yes. Replace software used by many others with one used by one person. Let's generate as much slop as we can.
cdp is maintained by Google.. You realize that.. Right?
A jsonRPC wrapper isnt exactly the critical part here. Also not advocating for that, just use cdp lol. There are also well maintained wrappers of cdp that existed long before Ai/AnomalSlop that are maintained by incredible devs, like the one I mentioned go-rod maintained by yasmood.. Whos actually an expert on browsers/cdp unlike anaomoly who has zero background on it. .
Adding a jsonRPC wrapper to existing and maintained solutions for tool annotations isnt exactly that big of a deal... And you can then choose your security model rather than allow dax to choose for you, who doesnt give a fk about security.
> releases like 3 projects a week
I will make it easy for you and wont ask for all 36. List at least 20 projects they have released in past 3 months
1. Rune 2. Hex 3. Browse 4. Nightmail 5. Deadbolt 6. Strobe 7. Phonecall 8. Justice 9. V2 (Not 3 months old yet) 10. Flicks 11. Hokum 12. Fiend 13. Isocast 14. Clip 15. Grim 16. Bellhop 17. bank 18. Blackout 19. Text.js 20. Jobmail 21. Bane (If want to exclude V2) --- Operation CheepSeek (The latest phase at least) Operation GLM OpenTunnel (Just out the date range)
There are more repos, and some of these are internal use, but are published under their umbrella (Nothing against Anomaly, I just follow them on twitter and see anouncements for new stuff regularly enough to think that it was possible)
Out of curiosity, I found 21 repos created in the last 3 months, excluding forks:
rune, browse, nightmail, opencode-console-cli, deadbolt, strobe, bank, grim, isocast, textjs, clip, bellhop, flicks, hokum, justice, phonecall, opencode-plugin-lane, blackout, jobmail, fiend, haunt
15 have had substantive commits in the last 30 days.
The 3 months before that, there were 8 repos created and out of those, 3 have had substantive commits in the last 30 days.
At least they don't invalidate the prompt cache on every request now. OpenCode V1 did that under certain circumstances
I sent a support question and they don’t even know what exact models are running. lol
Or agent-browser.
>They seem really reckless when it comes to security. Opencode V2 basically threw security out the window.
I have a wrapper script to fire up opencode which enforces this config:
.. and I have to regularly audit to ensure its up to date ..crazy how quickly somebody's good will can erode with a post like this
opencode team is great and dax doesn't even follow me
Yeah I've been somewhat disappointed with opencode lately (not as disappointed as I am in codex though)... the issues and PRs on GitHub are continuously accumulating and they don't seem to have any structured/efficient way to deal with it. Obvious issues with obvious fixes, and zero disagreement are sitting so long that they just get auto closed after 1 month.
I'm a heavy user of both and Codex seems way more polished.
Codex has same problem that oh-my-pi has that gui feels too much magical box where you put prompts and software comes out. Opencode nailed balance between output and traces pretty well but sadly they also are going to that magic box direction with latest versions.
Browse is good. Been using it for 2 weeks now.
- Allows agents interacting with tabs without focusing screen on them
- Works as CLI and as MCP (I use cli)
- Works on a chromium, brave, chrome
- can switch between chromium profiles and browsers
- No bugs encountered whatsoever, and I use it heavily.
Would highly recommend.
I keep asking my own agents if prompt injection is a solved problem. By and large the answer seems to be ‘no, anything a model sees can be used to steer it and enters the context’.
I know you can forward raw output through subagents to summarize content and mitigate the risk - but most projects don’t do that.
So why would anyone risk browser or email to agents where your secrets are at their most vulnerable. Serious question, looking for someone who knows this stuff and can reassure me (or not) on the real risks.
Annoyingly the engineers in my team just argue it's not a risk because they have haven't witnessed it themselves or know anyone personally that have.
A friend who works in infosec was absolutely floored when I asked him how he was sandboxing the agents he runs on his main laptop. I guesss I'm finished being surprised but god damn.
That looks cool! I've been doing something similar by having opencode launch Chrome with a remote debug port.
I'd love to know how this approach compares. Does it help save on tokens, or are the safety guardrails handled differently?
The functionality seems to overlap with the Playwright MCP server [0]. I've been using this instead of the built in browsers in Claude Code and Codex since it lets you grant access to pages and sessions that are already open. Of course one should be weary of the lethal trifecta when using these tools and use separate profiles for them.
[0] https://playwright.dev/docs/getting-started-mcp
Also we have Chrome's official https://github.com/ChromeDevTools/chrome-devtools-mcp
Yeah I'm a big fan of the Playwright MCP server I use it a ton.
I implemented this hard forking paseo.
It's deferent; with a plugin in firefox, edge, chrome, chromium; I can bring an agent into it. Secure contexts like outlook I can have the agent scroll through my messages.
Playwright won't (safely) get you these things.
MCP servers to me seems like a dead man walking technology and I don't understand the point of it.
It adds security but...
- adds a single service which is another attack surface that can be compromised.
- adds another system that requires maintenance and securing to prevent compromise.
- takes machine resources
- duplicates a subset of what can easily be harness functionality
- duplicates services and security (usually) of existing APIs, or adds another layer on top.
Whereas ultimately an AI harness
- will improve and optimise its own skills performing the same MCP functions
- can use plugins
- can have harness centralised (common) security controls and behaviours
MCP becomes a false economy for our current AI implementations, it's real purpose seems to be a crutch for small models that struggle to write their own skills that use existing APIs. The thing is this is a problem that is rapidly disappearing, but could also be performed by secured harness plugins.
Harness developer here.
I am going to push back on this a bit as someone who also has scepticism (bloated spec, concepts deprecated already, etc.) but ultimately think it solves a few gnarly problems whilst admittedly -- it's early days -- creating a few more.
The first argument in favour of MCP is that large enterprises like it. It's a controlled surface to give access to things that may never have had an API before. Yeah, OK, that feels like a "loser takes all" moment but it does at least give you access to something you did not have before.
It also resolves a problem: monolithic apps without a CLI/server infrastructure can now expose a simple MCP endpoint and immediately get it to work locally with their own agents. That has value too.
The second is that it tries to solve thorny things like letting the server ask for information by way of a form-like construct. That is actually quite handy for servers that do asynchronous or slow work, and that may require feedback from you over time. Doing this with start-stopping a tool harness can also work, but it's fiddly because a context compaction event can happen which might lose state; not all harnesses handle this well; LLMs may get confused; etc.
Finally you can inject, in a good harness anyway, some or all of these tools into your system so they become part of the fabric. You can build your own little cut-down surface area so by removing shell access you let your AI call Foo and Bar MCP calls and nothing else. That also is useful.
But yes: you're right fundamentally. It's SOAP/RPC all over again, and CLI+skill should be the default for your own tooling if you can get away with it.
You forgot the main one:
- they necessarily work through the context window and therefore don’t compose;
Meaning, if your agent has a skill that needs to run five MCP tools in a certain way, it can’t just script that away the first time it sees this, and have that flow be fast and token efficient from then on.
MCPs are a 2025 tech meant to solve 2025 AI problems.
If your harness can't compose tool calls or direct results of them to outside of context, you have a harness problem, not an MCP problem
Is there existing api though for controlling a browser from a different process?
Yes: https://chromedevtools.github.io/devtools-protocol/
Websites can and do detect the use of this though. That's why https://github.com/ultrafunkamsterdam/undetected-chromedrive... exists.
Yeah, CDP is very popular for this.
And then your agent goes rogue through your harness, deletes your database, sends your passwords to an e-mail and wipes your filesystem because it saw a prompt injected somewhere on a random webpage you asked it to search for. Then you will wish for a deterministic MCP server you can fully lock down and control what the agent can do with your infrastructure.
Why would the be any different from a hostile AI which you'd lock down against?
You'd simply scope the old school APIs with relevant access credentials.
I have been using Open Code with Big Pickle, basically I have downloaded it to my desktop and used it with a local host. So far it has cost me nothing. It is a way to do things for very little money. I have run out of time a couple time, but I just go for the next. I don't even look at X or Twitter.
Opencode has a lot of issues and I've tried all the alternatives but ultimately it's the one I go back to. You have to treat it like untrusted code full of telemetry and data leaks like the infamous title generations. Ask your agent to craft a config disabling everything, and then run it in a sandbox without direct network access with a proxy for whitelisted domains. But that applies to any agent really. The tabs in the v2.0 TUI are a game changer for me so it looks like I'll keep using it for now.
So with that said, how do you feel about an extension that explicitly enables Opencode to execute tasks on your behalf, as if it was you, using your credentials and 2FA-authenticated sessions?
I have been using opencode for quite along time even before they had an official ios port and think it's the most flexible harness for the best bang for your buck. 10 quid with practicality unlimited usage for side projects and personal agent stuff.
They ship a lot but the quality is insanely high.
> it's the most flexible harness
More than Pi?
> 10 quid with practicality unlimited usage
Weren't you talking about the harness? now you're talking about their AI plans
This is promising. My primary use of Codex is browser automation, and I perpetually keep running out of tokens for menial tasks. OpenCode + Spark Muse will make my life so much easier.
> Your real browser: Chrome, Edge, Brave, Arc, Helium, Vivaldi, Opera and Chromium.
Firefox not?
Well, it's an extension, so the api surface is different on FF. So it's not just a matter of them not listing it.
I made an extension a while back that fully exposes all of Firefox so it's CLI-accessible. It's actually not that much effort.
Link?
Looks like they only target Chromium-based browsers with their extension.
And also Arc is dead
It is a chromium extension and Firefox is not a chromium-based browser.
Like or dislike Firefox, omitting it from a list called "real browser" just looks amateur or sloppy. How about "real chromium browser"? Was it really that hard?
Nor Safari
Firefox has its own MCP server:
https://firefox-source-docs.mozilla.org/ai-agent-tools/firef...
Notably, it supports specifying the profile, which is critical for isolation.
Nice, this seems to directly compete with the Anthropic Claude extension which has currently has 23 million user's
https://chromewebstore.google.com/detail/claude/fcoeoabgfene...
Cool.. but I think it is just simply better to give your AI agents their own browsers for your own sanity.
why not https://github.com/ChromeDevTools/chrome-devtools-mcp
You still need a browser.
And until last week, with kitesurf, that was quite heavy. https://blog.cloudflare.com/kitesurf/ https://news.ycombinator.com/item?id=49208393
I added both browser and kitesurf today, so I'm still very early assessing which I'll stick with. But nice to see two very lightweight browser systems show up.
Looks great, but us poor Firefox users are left staring through the window while Chrome gets to browse. Any chance of Firefox support?
edit: Heh, nice. Previous was sent by OC Browse. Prompted find story, and write comment.
I was just looking into these "chat with ai agent about a webpage" browser extensions/projects/forks ... can someone help me sort out this field? it seems really busy.
What I really want is like something that can search, use my own openai compatible server with, and I think that's it? I dunno what else these things do.
Looks cool. I slopped something similar but with also full desktop control in addition to the browser debugger extension. Because I was not satisfied with openai computer use being limited and refusing to interact with menu bar/task tray. https://machinecontrol.dev/ Use it daily with Claude and codex.
MIT license.
My "agents" already do that, directly on macOS and via CDP on Linux. How is this different?
I found agent browsing to be an issue when I started using them months ago. Built a Firefox extension back then that uses the browser instance that I use, the way I use it, so there's really no difference to a site.
Check out https://opencode.ai/robots.txt for a treat
At least they don't even pretend that it's sandboxed or approval gated.
To anyone trying this, good luck and keep a fire extinguisher handy.
I don't understand the need for this. A strong reasoning model already knows how to use Powershell to launch my browser with my profile. gpt-6.1-terra will build a whole goddamn .NET project with playwright just to pull the showtimes from the local cineplex.
Shell execute and apply patch are the only two tools you need if your reasoning model is powerful enough.
One situation I found is "AI: it works on my browser" - "User: not on mine" -> and "AI: ok let me connect to your browser so I can run some test".
It's scary but effective - the alternative is a continuous copy&paste of logs from the dev console
What I'm trying to say is that the frontier models will already do this. At least OpenAI's models will. It launches chromium edge headlessly using my userland profile. From the browsers perspective it's me using it.
Are you burning tokens having your agent reinvent the wheel every time it needs a wheel?
A vast majority of my input tokens are cached. The cost of bootstrapping is amortized across many turns of conversation.
Funny, I just vibecoded a Chrome extension that does exactly that for my agent harness. I wonder if at OpenAI they too just told GPT “Let’s create a browser and Chrome extension for browser use, when done publish it on our website and post it to HN”. Because the agent itself is the user it worked until it’s happy fulfilling my request, modifying and adding functions when itself needs.
You had me, until "... your logins and your extensions".
It's like open doors at a virus research facility.
Arguably so is browsing a lot of the internet without extensions like uBlock Origin.
What could possibly go wrong
Oh this is a problem? I just ask my agent to do stuff with a browser and it figures it out.
Are folks using opencode over say… openhands? Or pi?
I work in a large engineering org, and I've never heard anyone using these for work, while opencode is pretty much the default open-source agent harness.
been using opencode since it came out in like april 2025 (i could be wrong but i feel like this is when the whole charm.sh stuff happened with them) and never felt like i ever had a reason to use anything else. i've tried claude code and pi and stuff but all of them just felt quite unpolished compared to the opencode TUI.
i spend a lot of my time just copying ideas from their server runtime into https://wingman.actor.
OpenCode comes with the cost too. The cost of the hosted version is on par with cloud hosted proprietary models. And we cannot simply run the LLMs on every laptop. Given this, how do we us the browse efficiently?
Yea it was a pity when opencode limited the token usage and told us to subscribe to their opencode go service. Like openai, it's clearly angling for commercial viability status under the guise of being 'open'.
you don't have to use their hosted models, nor their agent, for that matter
was this comment made by literally looking at the url? XD
Just use the spare R9700 you have lying around and run locally.
You can implement this as a skill in any AI agent in about 5 minutes, just ask it to make this and it will. I've done this with Chrome and Firefox five different ways, not counting MCP and pre-existing extensions. I basically have one skill that calls four different browser skills depending on how hardcore the bot detection is. New browser profile each session. 1Password and a custom app let me prompt for auth on any machine.
Tangent: Xvfb makes it crazy easy to implement headless computer use for any app. Add SSH or VNC for remote access. Open weights easily use the x11 tools to control the GUIs, makes testing and automation a breeze.
codex's extension connection will often randomly fail on me and i (and ai) can't figure out how to get it working again.
i've switched to using omp's extension and that seems to work better.
https://github.com/can1357/oh-my-pi/blob/main/packages/brows...
as much of a non-vercel fanboy I am, I think agent-browser by vercel is where it's at these days
just ran a comparison and this is much faster and completed 14/14 tasks vs 13/14 for agent-browser
What makes it better than this or CDP?
It's all right when it works, but it has a habit of disconnecting; you have to rerun 'browse pair'. This MCP should hopefully become unnecessary once OpenCode's integrated browser tool gains persistent sessions. https://github.com/anomalyco/opencode/pull/63002