Provider: LM Studio

LM Studio is a desktop app for downloading and running local models. It includes an OpenAI-compatible server, so VibeCody talks to it the same way it talks to a hosted provider — with a local base URL and, normally, no API key.

Use it when you want a GUI for browsing and swapping models, and you would rather not manage a Python inference stack.

Requirements

LM Studio installed
At least one model downloaded through its Discover tab

Start the server

In LM Studio: Developer tab → load a model → Start Server.

It listens on http://localhost:1234/v1 by default — the URL VibeCody assumes. The Developer tab shows the exact model id it will answer to; that string is what you pass as the model.

To confirm it is up:

curl http://localhost:1234/v1/models

Configure VibeCody

LM Studio accepts any bearer token, including none, so there is normally nothing to configure. All three routes work:

Nothing at all — with the server on its default port, vibecli --provider lmstudio works as-is.

Config file (~/.vibecli/config.toml):

[lmstudio]
api_url = "http://localhost:1234/v1"      # default
model = "qwen2.5-coder-7b-instruct"

Environment — LMSTUDIO_BASE_URL overrides the URL; LMSTUDIO_API_KEY exists for completeness and is rarely needed:

export LMSTUDIO_BASE_URL=http://localhost:1234/v1

Models

LM Studio serves whichever model you loaded in the app. The list below is what the pickers offer as a starting point, not a claim about your machine — every picker also accepts a typed-in id, and /v1/models on your running server is the authority.

Model ids are LM Studio’s own short names, which differ from the Hugging Face repo paths used by vLLM: qwen2.5-coder-7b-instruct, not Qwen/Qwen2.5-Coder-7B-Instruct.

Model	Notes
`qwen2.5-coder-7b-instruct`	Default. Good coding model on modest hardware
`qwen2.5-coder-14b-instruct`	Stronger; wants more RAM/VRAM
`meta-llama-3.1-8b-instruct`	General-purpose
`mistral-7b-instruct-v0.3`	Small and fast
`phi-3.5-mini-instruct`	Smallest of the set

Verify

vibecli --provider lmstudio "Explain the borrow checker in two sentences"

Troubleshooting

Connection refused — the server is not running. Loading a model in LM Studio is not enough; the Developer tab’s Start Server must be on.

404 on /chat/completions — the base URL is missing the /v1 suffix. VibeCody appends /chat/completions, so api_url must end in /v1.

“Model not found” — the id in your config does not match what is loaded. Copy the id from the Developer tab, or from curl http://localhost:1234/v1/models. The short-name convention is easy to get wrong if you are used to HF paths.

Replies are very slow — LM Studio falls back to CPU when a model does not fit in VRAM. Its own UI reports the split; a smaller model or a heavier quantisation usually fixes it.