MinniCode: A free AI coding agent that runs on OpenRouter models
MinniCode works in your project the way a developer would: it reads the code, makes edits, runs the tests and opens pull requests. It runs on free AI models from OpenRouter with your own free OpenRouter key, so there is nothing to pay.
Downloads Paused
$ minni MinniCode - the Open-LLM Coding Agent /help commands /settings model, theme !cmd shell /exit > add tests for sample.py and fix whatever fails read_file sample.py write_file tests/test_sample.py [Enter approves] run_tests pytest -q 2 failed edit_file sample.py run_tests pytest -q 5 passed changed: sample.py, tests/test_sample.py
Free to run on OpenRouter models
Runs on free OpenRouter models with your own key. Switch models any time.
You approve every step
Every file edit is shown to you first, and nothing changes until you accept it. Riskier actions, like running a command, deleting a file or pushing to GitHub, need a clear "yes" from you before they happen, so nothing risky goes through by accident.
Made for free models
It loads only the project notes a task needs, so a free model spends its attention on your code.
- Approvals
- Every edit is shown to you first. Running commands, deleting files and pushing need a clear yes from you.
- Memory
- A small set of notes stays loaded every turn, and your project’s MINNI.md file is read only when a task needs it: the same idea as MinniMemoryMCP.
- Models
- Free AI models from OpenRouter, reached through Minni’s servers on your own OpenRouter key. You can switch models each session.
- Pricing
- Free.
- Latest
- v0.6.0, released 2026-10-05.
How a turn works
You type a task. The agent reads and searches first, then shows you each edit to accept. Running commands, deleting files and pushing need a clear yes from you. It runs the tests before it says the work is done.
Where the model comes from
Free and open AI models from OpenRouter, reached through Minni’s servers on your own OpenRouter key; you can switch models each session. A free key is enough: MinniCode asks for it the first time you run it.
Talk to us before it gets expensive.
Send us what your agent loads each turn. We apply the interface and report tokens saved against answer quality.