If you want to use local models, it's more ergonomic than fussing with GGUFs or using LM Studio as a server host and setting up the link to an agent yourself. Although, the model selector is the same as with LM Studio itself which can be overwhelming if you don't know what to look for.
Built to work with lmstudio, one of the leading easy to use local model servers. LMStudio is the closest to plug-and-play without sacrificing play that I've seen; a harness that works well with it is nothing to sniff at. Its not earth shattering either.
I guess it lives or dies by the harness quality then - on open models run locally by plug and players and models that fit onto peoples laptops that is going to be quite the handicap to overcome.
I run lmstudio personally with a range of harnesses (open and closed) and can't say there is that much of a leap to getting everything talking https://lmstudio.ai/docs/integrations
Maybe they optimized it really hard for small models. That would be impressive, but hardly a good selling argument. I want a harness that can do both and OpenCode looks just fine for that purpose (and it also has everything else I need).
To me this looks like another case of bundling things that shouldn’t be bundled (the harness with the UI) making both worse off because you can’t individually focus on each component. They’ve done this before, bundling a decent UI with decent inference for IMO no good reason, combining the downsides of each instead of letting people mix+match.
Wouldn't most opensource harnesses work with lmstudio? I assume it has an "openai" style chat API like every other model provider? What's special about it vs langchain deep agents or pi or pydantic-ai?
Yes. I don’t see it either. It looks like a competent app (converging on the same principles as others) but what they are advertising as differentiators simply isn’t relevant to its purpose.
built to work with an OpenAI API compatible endpoint, just like any other harness...
and if someone can't figure out how to write down an address it's very likely they also can't figure out how to make local models not suck for coding, and would likely switch back to codex/cc after 15 minutes anyways.
Ultimately, the onus at every VC backed local LLM startup is to launch a cloud based offering, because that's the only potential path in sight for venture scale returns.
For now, it seems that direction that lm studio is taking for enterprise market is "local ai deployment support". They recently launched "lm link" which basically uses tailscale to create e2ee connections between computers running lm studio where you are logged in. Granted one can also setup tailscale or their own vpn themselves and use llama-server, but I guess it is simpler to provide it out of the box. In any case I am not sure pivoting from running local models to "cloud offering" (as in providing llm inference at their severs) is a sensible choice granted there is already competition in that space and they have no leverage there. The highest expected path imo would be to be bought by a company that makes (esp open weights) llms and has a similar business plan around enterprise contracts with local deployments.
> In any case I am not sure pivoting from running local models to "cloud offering" (as in providing llm inference at their severs) is a sensible choice granted there is already competition in that space and they have no leverage there.
I agree. Incidentally, this is exactly what ollama is doing too.
I can definitely see a world where you run stuff local first, for all the reasons we know. Sometimes, you are going to want more powerful models, faster, you’re travelling, etc. You might only use these 5 to 10% of the time, but I’m guessing that’s the market they want.
Yup, it's the main reason I don't use LM studio more. I only use it to try out new models/quants, then use llama.cpp directly to host them. LM Studio also doesn't do stuff like audio input and often has bugs that pure llama.cpp doesn't so it can be a net negative for certain use cases.
I am aware of it, and I dabble with Unsloth Studio and use the llama-server approach.
I would obviously prefer an open source, open weights stack.
But I guess a paradox is that as long as there are open source options I could use, a solid agentic environment that I can use with my own open weights is something I might pay for, in a similar sort of way to paying for a Mac when I could use only Linux.
If someone wanted to make their entire income from, say, making the BBEdit of LLM harnesses, that would be a viable strategy. Sooner or later people need to make an income somewhere. My own feeling is that Apple should acquire LM Studio, but if they said "this is $X per year" I might consider it, given the attention to detail.
I seriously don't understand the apple glazing. They are pretty much completely out of the picture and are one of the most anti consumer anti open-source companies. Yet developers STILL hope for Apple acquisitions and being saved... Or something like that...
We have llama.cpp, vLLM, opencode, hermes and countless other mature great open source tools that don't require a anti-consumer company's hands in the jar.
This kind of thing just makes me think Apple will get to a point where they have good enough local models and good enough harnesses for doing things, and most normal people will just use them… Does the LLM become another interface to computing?
i believe that for most people on the street, for most tasks, a Chat GPT 3.5 era LLM is sufficient enough. sprinkle in tool calling and other things, and that becomes enough. if you can prioritize that level of a model on-device (baking it in etc), then you can bifurcate AI users between those unwilling to pay and those who are willing to pay A LOT for frontier model performance.
I have thought this for a while. Computing 1.0 meant that we needed to learn the computer’s language to interact with the computer fully. Computing 2.0 is that now the computer has learned our language instead.
This question hinges on whether model advancement plateaus enough for machine sized models to compare to frontier performance. If it does, the answer is yes. If it doesn’t, the answer is no
More likely, it's going to be whether frontier models advance enough that most people would be willing to pay for them. Right now they don't, but a model you can run locally for free on hardware you already own is very compelling because, while they're not as good as Frontier Models, they're still pretty good.
Tools like Opencode demonstrate that when you box them in tightly enough they can actually be pretty competent.
I disagree. The point of the frontier models is to do everything as well as possible ("AGI" race or whatever) but smaller models with some RL are going to be the clear winner for a ton of use cases. Think about all the use cases for LLMs that would never be economical at frontier inference costs, and in no way need it. You don't need or even want a phd polymath helping you with small productivity tasks that most people use computers for every day. It's often overwrought and annoying. I don't even really like the frontier models for coding for this reason. They're constantly blowing up scope and you have to fight it constantly.
Neural machines were always going to be an alternative computing paradigm to von Neumann machines. Had it not been for Minsky we would arguably have gotten to a point where they're useful sooner. But why do you say that as if it's a small thing?
Why wait? People are already doing their work on OpenAI and Anthropic's servers, Apple Intelligence servers could quickly subsume any "local" model work that you want to do.
That way everyone has access, even with older devices, and it's a subscription! Then Apple can tie their APIs into the ecosystem you love at a flat cost you can afford. No need to support local model integration in the first place, problem solved.
based on their apple intelligence demos they are optimizing their products for their core demo of 55-95 year old boomers who talk out loud to think and read every page of the nytimes. you are miles away from the US consumer product experience here.
It does! We negotiated ZDR with our providers. We consider that a condition to make things available to our users: in this case cloud inference and web search/extract.
Why would I use this over any other harness? I suppose they're wrapping it all up in a nice package for enterprise, especially ones that want to control their LLM usage for cost and data security compared to the cloud frontier models.
There’s actually not a lot of good harnesses that aren’t slopped together python or JS codebases, which are actually model-agnostic and don’t do some really silly things like bloating the context, too much compaction etc.
There is no way I’m running a python or JS agent which is probably vibe coded, it’s just too much risk from a security and supply chain standpoint.
I don’t vibe code though, to the extent I actually use agents to generate code it all gets reviewed and it’s essentially exactly what I would have written myself. It’s mostly things like mechanical refactors, boilerplate generation, etc.
https://github.com/pjlsergeant/byre can add small but very convenient layer around whichever agent you want, initially trapping it in a specific folder, but you can easily mount more, firewall it, bring in MCPs and skills, or whatever. It's unambiguously AI-assisted, but a very very great deal of thought and work has gone in to the design of it. Ask your favourite agent to review the code if you're skeptical.
Hey everyone! Yagil the founder of LM Studio here. If you want to take Bionic for a spin with GLM 5.2 / Kimi K2.6 / Kimi Coder K2.7, email your lmstudio.ai username to hn-jul16@lmstudio.ai and I'll load your account with some credits!
Try it out for coding (in a "Code" project) and document creation / manipulation (in a "Work" project). In Work projects we have automatic checkpointing for every change the agent makes. Would love to hear your feedback.
Yagil, thanks for the credits! Lots of fun to try running this locally. I don't understand the negativity in some of the other comments.
This is one of the better agent harnesses I've seen for inspecting reasoning chains, which is super useful for me. Sometimes reading the reasoning is better for my needs than reading the response. I appreciate how transparent that is in contrast with Claude Code/Codex/etc.
One question — I saw you mentioned negotiating ZDR with your "providers" — are you hosting the models yourselves, or is another entity hosting them? If another entity, which group?
I have never previously tried a agentic harness for local models yet, but I really love LM Studio so I gave Bionic a shot immediately after reading this!
It works great! I use Codex as my main agent, and the UI looks similar enough that it's familiar and simple to get started. I just pointed it to my LM Studio models library, ran Qwen3.6 35B, and the results are exactly what I would hope for.
I did notice some rough edges that might be worth improving, however:
- Current working directory is not the clearest on the main page of the app. It shows the project name, but is missing the prominent working directory label like Codex has.
- The model seems to load when you hit Enter, but it shows "Working" instead of of "Loading model".
- There doesn't seem to be a way to pre load the model, it seems like you have to send it something to load the model.
- I don't see a way to easily unload the model like the eject button in LM Studio without quitting the app
- I pointed it to a directory called "GitHub & Projects" and it somehow ended up making a new folder called "GitHub & Projects". Yes, I know the name is weird but it shouldn't have done that.
63 comments
[ 0.19 ms ] story [ 14.1 ms ] threadI run lmstudio personally with a range of harnesses (open and closed) and can't say there is that much of a leap to getting everything talking https://lmstudio.ai/docs/integrations
To me this looks like another case of bundling things that shouldn’t be bundled (the harness with the UI) making both worse off because you can’t individually focus on each component. They’ve done this before, bundling a decent UI with decent inference for IMO no good reason, combining the downsides of each instead of letting people mix+match.
and if someone can't figure out how to write down an address it's very likely they also can't figure out how to make local models not suck for coding, and would likely switch back to codex/cc after 15 minutes anyways.
I agree. Incidentally, this is exactly what ollama is doing too.
Since most people are unaware of this fact.
I don't think we need closed-source developer tools, especially ones where they might restrict access if they decide to start charging for them later.
I would obviously prefer an open source, open weights stack.
But I guess a paradox is that as long as there are open source options I could use, a solid agentic environment that I can use with my own open weights is something I might pay for, in a similar sort of way to paying for a Mac when I could use only Linux.
If someone wanted to make their entire income from, say, making the BBEdit of LLM harnesses, that would be a viable strategy. Sooner or later people need to make an income somewhere. My own feeling is that Apple should acquire LM Studio, but if they said "this is $X per year" I might consider it, given the attention to detail.
i believe that for most people on the street, for most tasks, a Chat GPT 3.5 era LLM is sufficient enough. sprinkle in tool calling and other things, and that becomes enough. if you can prioritize that level of a model on-device (baking it in etc), then you can bifurcate AI users between those unwilling to pay and those who are willing to pay A LOT for frontier model performance.
Tools like Opencode demonstrate that when you box them in tightly enough they can actually be pretty competent.
That way everyone has access, even with older devices, and it's a subscription! Then Apple can tie their APIs into the ecosystem you love at a flat cost you can afford. No need to support local model integration in the first place, problem solved.
Please don't slur us older GenX as boomers.
It already is.
(I’m the founder of LM Studio)
> use the largest frontier open source models through LM Studio Secure Cloud
There is no way I’m running a python or JS agent which is probably vibe coded, it’s just too much risk from a security and supply chain standpoint.
Try it out for coding (in a "Code" project) and document creation / manipulation (in a "Work" project). In Work projects we have automatic checkpointing for every change the agent makes. Would love to hear your feedback.
This is one of the better agent harnesses I've seen for inspecting reasoning chains, which is super useful for me. Sometimes reading the reasoning is better for my needs than reading the response. I appreciate how transparent that is in contrast with Claude Code/Codex/etc.
One question — I saw you mentioned negotiating ZDR with your "providers" — are you hosting the models yourselves, or is another entity hosting them? If another entity, which group?
Nowhere near ready for production
It works great! I use Codex as my main agent, and the UI looks similar enough that it's familiar and simple to get started. I just pointed it to my LM Studio models library, ran Qwen3.6 35B, and the results are exactly what I would hope for.
I did notice some rough edges that might be worth improving, however: - Current working directory is not the clearest on the main page of the app. It shows the project name, but is missing the prominent working directory label like Codex has. - The model seems to load when you hit Enter, but it shows "Working" instead of of "Loading model". - There doesn't seem to be a way to pre load the model, it seems like you have to send it something to load the model. - I don't see a way to easily unload the model like the eject button in LM Studio without quitting the app - I pointed it to a directory called "GitHub & Projects" and it somehow ended up making a new folder called "GitHub & Projects". Yes, I know the name is weird but it shouldn't have done that.
* locked to a single dir, so no system wide access.
* no local web search, can be fixed ddg or local mcp.
* no ssh, I want to have it ssh into my server and do the work.
* doesnt show the model being loaded, needs a bar/% counter.
* Can you drag/drop documents in the work dirs, or only + add them?
I love lm-studio, so cant wait to see how this goes. For local I normally use opencode + lmstudio.