How to actually have agent write code?

[capppybara]

Hi,

I am running deepseek r1 14B locally on 32gb ram macbook pro, and chat + deep research all seems to work ok. However I can't for the life of me get the agent to actually write code, it will always just give instructions and code snippets but won't touch files, and half the time it will give me this error Endpoint offline even though the chat works fine.

[reisnobre]

I don't know if this project is meant write code like Claude or Open code, I might be wrong, but I think Odysseus is more a way to interact with local LLM's in a mode structured way for handling tasks, processing text, research and such, but I don't think it's an IDE, but again, I might be wrong

[RlySkiz]

Why does it let you filter for coding specific agents then

[reisnobre]

Don't know exactly, but maybe it inherited that from hugging face? Also, I'm using the project right now to analyze the project code, so probably I should be using a coding model for that, but I don't care enough atm.

[sacb0y]

It's a feasible usecase with an MCP but the problem is there's no documentation on setting that up in Odysseus at the moment, so unless you're knowledgeable (or ask github copilot) you're at a loss for how to set this up.

[capppybara]

I'm pretty sure agent mode's purpose is to code when the task demands it. But good point on the MCP I'll look into if I can unblock it through that.

[CoolJohn-lab]

It definitely should be able to write code. I've had it write a few python scripts. Some of the agentic tool calling stuff is screwy though. There's no dedicated coder function or any kind of IDE so it's not exactly designed to write code per se, but it can be made to in a pinch.

It seems like the best way to use your Odysseus hosted models to code right now would be to run it alongside something like Claude Code, OpenCode or Cline for VS Code. (Or all three, why not they're all free 😌) and you point those tools at your local coder model that Oddy hosts for you in llama.cpp or whatever.

You can then chat with Odysseus about your repo, have it help you plan and keep track of changes and suggest next steps etc. then copy-paste its advice into your coding harness and do the actual work there. With some work you can set it up with a GitHub MCP so it can check the code and give you feedback on how good your work in the IDE was...

[RainOrigami]

I think I have misunderstood the point of this repo as well. I wished it were like claude where I can tell it to fix an issue with my system and off it goes reading configs and such.

Even a simple thing like having it code up a rust app that shows the current date and time it fails horribly at (Qwen3-Coder-30B-A3B-Instruct-AWQ) getting stuck in a loop describing the code it wrote but failing to install the tools required to compile it (because it runs in a docker container?).

I'm really searching for a claude cli replacement where I can run the models on my gpu instead but alas, this one wasn't it either.

Cool stuff nonetheless.

[CoolJohn-lab]

Yeah so the tools getting caught in a loop thing is a bug. There's hundreds of bug fixes awaiting approval, it will get a lot better at coding.

But I think you have missed a big feature of Claude Code. It doesn't have to use Claude as its LLM, you can point Claude Code at any model, so you can carry on using that. Just point it at a locally hosted coder model that you run in Odyssey. Thats basically the entire point of Qwen I think, to replace the Claude in Claude Code.

Or like the user above suggests, get OpenCode and point that at your local model.

[zurie]

You need to download OL

If you have a model in huggingface (Odysseus) . you need to convert it to Ollama first. (Make sure you set up proper context tokens)

For example, I downloaded qwen3-coder-next (48gigs) for my 96gig M3 Ultra Mac. (downloaded the model in Odysseus)

after I converted it to ollama, and now... I can type.

ollama launch claude

from there it lets me pick any model I want.

As long as you start ollama launch claude from the directory you want to work in.

Now you can tell claude something like "Create a markdown file" or Make a svg of a robot. or Install the latest angular21

it will do that for you. I use it to do agentic coding and i've been throwing enterprise level tasks at it for the past few weeks and its as good as Claude 4.5 which was good enough for big tech last year ;)

[capppybara]

You need to download OL

If you have a model in huggingface (Odysseus) . you need to convert it to Ollama first. (Make sure you set up proper context tokens)

For example, I downloaded qwen3-coder-next (48gigs) for my 96gig M3 Ultra Mac. (downloaded the model in Odysseus)

after I converted it to ollama, and now... I can type.

ollama launch claude

from there it lets me pick any model I want.

As long as you start ollama launch claude from the directory you want to work in.

Now you can tell claude something like "Create a markdown file" or Make a svg of a robot. or Install the latest angular21

it will do that for you. I use it to do agentic coding and i've been throwing enterprise level tasks at it for the past few weeks and its as good as Claude 4.5 which was good enough for big tech last year ;)

Yeah I mean I was hoping Odysseus could've been the all in 1 app, but if I have to use ollama I might as well go back to my old setup.

[CoolJohn-lab]

You don't have to use ollama, and you don't have to launch claude code from ollama either.

Ollama runs at about 60% of the speed of vLLM or Sglang so you gain a lot of performance hosting in Odyssey (when it works) and you can have Claude Code and Oddy both loaded in an IDE for the best of all three worlds.

You guys just need some imagination.

[zurie]

1. its not 60% faster... its marginally faster depending on hardware... NVIDIA may favor vLLM or Sglang over ollama, that is true, but its not 60%... don't scare the poor guy.. https://arxiv.org/abs/2511.05502

2. ollama is a convience wrapper for launching claude, if you are asking how to write code / agents then probably ollama would be easier than doing the manual connections. but you absolutely can just launch claude by typing claude and if you point to your anthropic url / key / etc then yes, it will use your models stored on odyssesus.

[vdmkenny]

Good question, and you're close, it just needs two things turned on, plus some context on where this is headed.

1. Use Agent mode, not plain chat. In plain chat the model only talks, it'll describe code and paste snippets but never touch files. The file-editing tools (read_file, write_file, edit_file, bash, etc.) only run when Agent mode is enabled (the agent toggle in the chat input). That's almost certainly why you're getting snippets instead of edits.

2. Select a workspace. Use the Select workspace picker (the "+" / overflow menu in the input bar) and point it at your project folder. That becomes the agent's working directory: file/code tools are confined to it and bash/python run there. Without a workspace the agent has no project to act in.

Where this is going: a capable, local-first coding agent is very much a goal for the project, not an afterthought. It's early though, and two things gate it today: (a) reliability leans heavily on the model, smaller local ones (deepseek-r1 14B, Qwen3-Coder-30B) often loop or "describe what they wrote" without emitting a real tool call, exactly what you're seeing, and a stronger model helps a lot; and (b) there are known bugs in the tools themselves, especially on Windows (some assume a Unix environment), which we're actively fixing. So the agent already reads configs and edits files, and that path is being pushed hard, just expect rough edges for now and that model + platform matter.

On Endpoint offline while chat works: that's a separate issue, the agent can resolve a different endpoint (utility/task) than your chat endpoint, so check that endpoint is reachable. If it persists with Agent mode + a workspace set, open a separate issue with your endpoint config and we'll dig in.

One more thing: all of the above (Agent mode, the workspace picker, the recent tool work) currently lives on the latest dev branch, not main. main is only updated at releases, so if you are on main you will not see these yet, switch to dev to try them.

[capppybara]

@vdmkenny what's the fix for macbook? how do I allow it to leave the sandbox?

[vdmkenny]

There is no sandbox in Odysseus right now, so I'm not really sure what to tell you. For me it works fine on native macOS. I'll dig a little and see if I can replicate it. Could you show me exactly what isn't working?

[vdmkenny]

Ah, maybe it's a macOS security thing. grant Full Disk Access to whatever launches Odysseus (Terminal/iTerm, or the Homebrew python binary /opt/homebrew/.../python3.11) in System Settings → Privacy & Security → Full Disk Access, then restart start-macos.sh. Or keep projects in an unprotected dir like ~/dev.

[capppybara]

I asked it to create a directory in ~/src/ and then make a hello_world.py file. It keeps creating the directory and file inside ~/src/odysseus/data/src/.

It also shares the thought process:

I see the issue. The write_file tool is failing because paths outside the sandboxed environment are not allowed. Let me help you create the folder and file within the allowed environment.

EDIT: let me try the access

[capppybara]

ok fixed, the access + adding the directory in the workspace made the hello world creation successful. thanks for the help!

[vdmkenny]

Quick follow-up on the Endpoint offline part: we just merged a fix for that on dev (#3205).

The cause was the reachability probe: Odysseus pinged the bare base URL to decide if an endpoint was online, but some OpenAI-compatible servers (llama-swap and similar, especially outside Docker) return 404 on the base /v1 while serving /v1/models fine. So the endpoint got flagged offline even though inference worked, which matches your "chat works but agent says offline" symptom. The probe now falls back to /models on a non-auth 404 before giving up.

Pull the latest dev and that side of it should clear up. If you still see Endpoint offline after updating (with Agent mode on and a workspace set), open a separate issue with your endpoint config and we'll dig in.

[capppybara]

I think this was the real answer to my specific problem. I was already on agent mode but it would constantly fail with this error. But for the general public the other answer is better.