open source · MIT · v0.1

A local coding agent,
straight from your terminal.

It reads, writes, runs commands and browses the web, for your self-hosted models. Four modes: it codes for you, explains its choices, guides you step by step, or reviews your code. Nothing leaves your machine.

Get started in 30s → Try it online ↓ GitHub
↳ works with Ollama ·llama.cpp ·LM Studio ·any OpenAI-compatible API
~/projects/chat · mini-claude
● Welcome to mini-claude


Built by
Hugo Stawiarski

model    llama3.2:3b
site     miniclaude.fr

Commands
/help     list every command
/model    switch or list models
/install  install locally


› 

enter send  ·  /help help  ·  /clear reset

↑ live demo powered by Claude · private and unlimited locally.

Why Mini-claude

A full coding agent, without leaving your terminal.

A coding agent that reads, writes, runs and browses the web, from your machine. Nothing goes to the cloud.

A real agent, not just a chat
It calls tools to get work done: read your files, edit them, run commands, search the web. An agentic loop, running locally.
Four modes, on the fly
auto just does it, explain codes and tells you why, guided refuses to code and asks the right question, review reads your code and flags what matters. /mode switches without restarting.
It browses from your machine
Mini-claude drives your own Chrome, searches and reads pages for you. No account to share, no remote session: your browser, your IP, your rules.
Local-first by design
Zero round-trips to the cloud. Your conversations never leave your machine. No telemetry. No accounts.
Diff before acting, undo after
Every state-changing action asks permission and shows a diff. In a git repo, each edit is auto-committed, /undo rewinds if the model goes off the rails.
Works with anything
Targets the OpenAI-compatible HTTP API. Ollama, llama.cpp, LM Studio, vLLM, all plug-and-play.

Install

Up and running in 30 seconds.

Make sure Ollama is running locally, pull a tool-calling model, then launch Mini-claude.

ollama pull qwen2.5-coder:7b
git clone https://github.com/hugostarte/Mini-Claude.git
cd Mini-Claude && go run ./cmd/tui

Prefer Docker? docker compose run tui works too.

How it works

Three beats. One loop.

01 · Ask
You speak, plainly

Describe the task in your own words. The model decides which tool to call, read a file, edit it, run a command, open a page.

02 · Tools
The agent acts

Mini-claude runs the tool on your machine and feeds the result back to the model. State-changing actions ask for your go-ahead first, diff included.

03 · Inference
Ollama (or any OpenAI API)

Runs locally. Mini-claude only speaks HTTP. Swap the backend, swap the model, same client.

FAQ

The questions we hear most.

Does my data or code leave my machine?
No. Mini-claude runs entirely locally and only talks to the model you host yourself. Your files, prompts and conversations never leave your machine. No telemetry, no account.
Which models and servers are supported?
Any backend that exposes an OpenAI-compatible HTTP API: Ollama, llama.cpp, LM Studio, vLLM, or your own server. Pick a model that supports tool calling, such as qwen2.5-coder.
Is Mini-claude free and open source?
Yes. The project is open source under the MIT license, free, and available on GitHub. You can use, modify and redistribute it freely.
Which operating systems does it run on?
macOS, Linux and Windows. It's a Go binary that runs in your terminal; a Docker launch is also available.
How is this different from a plain chat?
It's an agent: it calls tools to read and edit your files, run commands and browse the web. Every state-changing action shows a diff and asks for permission, and `/undo` rewinds the last edit inside a git repo.

On the blog

Field notes, plain English.

Tutorials, comparisons and lessons learned about local coding agents.

All posts →

Built by Hugo Stawiarski

Open source · MIT · github.com/hugostarte/Mini-Claude