https://github.com/andrewchambers/minicoder
A minimalist AI terminal agent and assistant for unix systems
https://github.com/andrewchambers/minicoder
agent ai cli
Last synced: 3 months ago
JSON representation
A minimalist AI terminal agent and assistant for unix systems
- Host: GitHub
- URL: https://github.com/andrewchambers/minicoder
- Owner: andrewchambers
- License: isc
- Created: 2025-07-21T09:13:17.000Z (11 months ago)
- Default Branch: master
- Last Pushed: 2025-08-06T23:20:51.000Z (11 months ago)
- Last Synced: 2025-11-30T03:45:55.843Z (7 months ago)
- Topics: agent, ai, cli
- Language: C
- Homepage: https://andrewchambers.github.io/minicoder/
- Size: 2.06 MB
- Stars: 8
- Watchers: 0
- Forks: 1
- Open Issues: 5
-
Metadata Files:
- Readme: README.md
- Funding: .github/FUNDING.yml
- License: LICENSE
- Support: support/build-with-cosmo.sh
Awesome Lists containing this project
README
# minicoder
**Website**: [https://andrewchambers.github.io/minicoder/](https://andrewchambers.github.io/minicoder/)
**Downloads**: [itch.io/minicoder](https://andrewchambers.itch.io/minicoder)
**Discord**: [Join our community](https://discord.gg/cZk3yxBxRS)
**⚠️ Alpha version** - Be an early tester and supporter!
`minicoder` is a command line tool that uses ai to assist the user at
any task you can do from the command line on unix systems.
It is designed around a few core ideas:
- Follow good cli and unix design practices.
- Keep the tool as simple as possible, relying on AI advances to improve.
## Examples
```
$ minicoder --files '*.c' 'Can you fix the compile errors in this project?'
```
```
$ minicoder --model openai/o3 'Can you explain the files in the current directory?'
```
## Support the Project
minicoder is free and open source software. You can build it yourself from source, or download pre-built statically linked binaries from [itch.io](https://andrewchambers.itch.io/minicoder) with pay-what-you-like pricing.
Your support helps maintain and improve this tool for everyone.
## Community
Join our Discord community to get help, share ideas, and discuss minicoder development:
**Discord**: [Join the minicoder Discord](https://discord.gg/cZk3yxBxRS)
## Quick Start
```bash
# Set up your API key (OpenRouter or OpenAI)
export OPENROUTER_API_KEY="your-key" # or OPENAI_API_KEY
# Basic usage
minicoder "Fix the bug in main.c"
# Choose a specific model
minicoder --model deepseek "Optimize this function"
# Include specific files
minicoder --files 'src/*.js tests/*.js' "Add unit tests"
```
## User manuals
- [minicoder(1)](https://andrewchambers.github.io/minicoder/minicoder.1.html) - Main manual page
- [minicoder-model-config(5)](https://andrewchambers.github.io/minicoder/minicoder-model-config.5.html) - Model configuration documentation
### Use your own AI models
See [minicoder-model-config(5)](https://andrewchambers.github.io/minicoder/minicoder-model-config.5.html) for details on configuring custom models.
### Sandboxing
Sandboxing is a work in progress.
We aim to provide a simplified interface to limit
the assistant to specific directories and alsto to disable internet access.
## Building
### Prerequisites
- C compiler (gcc, clang, etc.)
- libcurl with development headers
- Standard POSIX build tools (make)
### Build from source
```bash
git clone https://github.com/andrewchambers/minicoder.git
cd minicoder
make
sudo make install PREFIX=/usr/local
```
The build process is straightforward - just run `make`. The Makefile will compile minicoder with your system's libcurl.
## Design notes
### Coding conventions
- The software is written in C to keep it resource light and portable.
- We use a tiny conservative garbage collector to prevent many of C's pitfalls.
- This project is often developed with AI assistance.
### Agent loop
- We try to make the agent loop is as simple as possible.
- The model is asked to keep track of its own plans and todos across iterations.
- The model is asked to 'focus' on files as it sees fit (add/remove from the context).
- The model performs actions in the world by generating shell scripts which are run, the output of which is returned to the model.
- There is no model/provider specific tool call convention.
- There is no infitely growing context, instead the model is fed its own compacted summary each iteration.