A local AI coding agent on LM Studio

Not just a chat — an agent

A model in LM Studio has no tools of its own — it can only call functions that were described to it. VibeCoder gives it six tools and runs them itself:

• Read, Glob, Grep — read a file, find files by name and text in the code;
• Write, Edit — create a file or precisely change a fragment;
• Bash — run a command in the project folder: tests, builds, git.

An agent turn is a loop: the model asks for a tool, the app runs it and returns the result, until the answer is ready. Every call is visible in the feed as a line.

Under your control

• Permissions. The model may read and search silently, but editing a file or running a command needs your approval — with the same banner as cloud agents. There is an “allow everything” mode and pause too.
• The boundary is the project folder. Every path is checked: the model cannot step outside the project, not even by accident.
• Long commands are killed. A command that does not finish in two minutes (say, an accidentally started server) is stopped together with all its child processes.

Set up in a minute

1. In LM Studio load a model and start the server (Developer tab → Start Server).
2. In VibeCoder open “Settings → Local model” and enter the gateway address — http://localhost:1234/v1 by default.
3. Press “Check”: the app shows which models are loaded. The model field can stay empty — then whatever is loaded right now is used.
4. Create a chat of the “LM Studio” kind in your project.

Which model fits

You need a model that supports function calling (tool use): for example the Qwen, GLM, Llama 3.1+ and Mistral families. LM Studio shows this capability on the model card.

If the model cannot call functions, nothing breaks: the app asks it again without tools and says in the chat why files are invisible to it.

Why local

• code and conversations never leave the computer — fine for closed projects;
• no subscriptions and no limits — the model runs as much as you need;
• works offline.