auditlm

A dead-simple, self-hostable code review bot with Forgejo integration. Sometimes you don't have a human to review your changes, and sometimes you don't want to give Microsoft more training data for Copilot. You can think of auditlm as an open-source and self-hosted CodeRabbit. I've using it to assist with its own development, so you can see it in action on some of the recent PRs.

What it does

auditlm is an AI-powered code review assistant that: - Monitors your Forgejo instance for pull requests that mention your bot - Sets up isolated container environments for safe code analysis - Uses local LLM models (like llama.cpp with GLM-4.5-Air) to review your code - Posts detailed reviews directly to your pull requests

Features

• Self-hosted: Run everything on your own infrastructure if you want!
• Privacy-focused: No data sent to third-party services unless you set it up with a cloud provider
• Flexible: Works with any OpenAI-compatible LLM endpoint
• Isolated: Each review runs in a clean container environment
• Language-agnostic: Analyze code in any language by providing a custom docker image

Quick Start

Prerequisites

• Docker or Podman for container management
• A self-hosted Forgejo instance
• An LLM endpoint (local llama.cpp setup recommended)
• Rust toolchain for building from source

Setting up Forgejo

First, set up a Forgejo instance using Docker Compose. You can do this with our example but I'd personally recommend deploying it with kubernetes instead if you plan on using it for more than just testing.

DOCKER_HOST="unix:///run/user/$(id -u)/podman/podman.sock" docker-compose up -d

Configure your bot

1. Create a new user in Forgejo called clanker (or whatever you prefer)
2. Generate an API token for this user
3. Export the token as an environment variable so that the auditlm binary can masquerade as your bot user:

export FORGEJO_TOKEN="your-api-token-here"

Set up your LLM

Configure llama.cpp with your preferred model. GLM-4.5-Air works well after applying this patch to llama.cpp and using the jinja template in that PR minus the trailing %. Note the base URL of your LLM, I gave mine a custom port to keep 8080 free, so for me it's http://127.0.0.1:11800.

Run auditlm

Clone this repository, navigate to the auditlm directory, and tweak the following command to your liking by modifying the docker socket and analysis image if necessary, then run it:

cargo run -- forgejo \
  --model 'GLM-4.5-Air-Q4_K_M-00001-of-00002.gguf' \
  --socket '/run/user/1000/podman/podman.sock' \
  --base-url 'http://127.0.0.1:11800' \
  --forgejo-url 'https://git.lunacy.casa' \
  --image 'rust:1-trixie'

auditlm will then listen for @mentions of the clanker bot in pull requests and provide a review in each case. Follow-up reviews will address changes since the last review.

Configuration

Docker Images

The example above uses rust:1-trixie for analysis environments, which won't work for all projects. Adjust the --image parameter based on your project's requirements.

Network Considerations

This is a toy example. While you can keep your forgejo instance running on localhost, you should consider making it public, because if the agent tries to connect to localhost inside the analysis container via e.g. curl, that request will not be routed to your Forgejo and the agent may get confused. Most of the agent's tools will still work, however.

Development

Building from Source

git clone https://github.com/your-repo/auditlm.git
cd auditlm
cargo build --release

Project Structure

• auditlm/src/commands/ - Command implementations (forgejo, git)
• auditlm/src/tools/ - AI agent tools
• auditlm/src/container.rs - Container management
• auditlm/prompts/ - AI prompts for different scenarios

License

This project is licensed under the AGPLv3 - see the LICENSE.md file for details.