Hello HN. I built a self-hosted code review agent. It is designed for homelab users, teams in closed network environments, and infrastructure teams running on-premises who need a lightweight and easy to deploy review tool.
I wanted something that could run in diverse environments like local LLMs and closed networks, and on self-hosted infrastructure like GitLab and Forgejo.
Works with GitLab, Forgejo, and GitHub
Bring your own LLM (Local or external or whatever, any OpenAI compatible endpoint, or Claude API)
Reviews pull requests and posts inline comments
Replies to users on pull request threads
Replies to issues
I ran a Martian benchmark with mimo v2.5 pro and placed 7th out of more than 20 code review tools (Precision 0.351, Recall 0.545, F1 0.427). I have not yet tested stronger models. I'll update with results when I do.
Built with Bun, SQLite, and SvelteKit on the dashboard. Everything runs in a single container and deploys in minutes.
I put a lot of time into this and would love for you to try it. Setup takes about 3 minutes only. You can use it for review on your current projects
Feedback, feature requests, and technical questions are all welcome.
I have couple of true (and therefore more expensive) CO2 sensors around the apartment, all of them Adafruit SCD-40, each connected to its own ESP32 which broadcasts the metrics (including temperature and humidity) to MQTT broker which feeds into Prometheus. A Grafana instance running on a Raspberry Pi can then post an alert which goes into PagerDuty and also to a small speaker connected to the RPi to play a distinct tune when the CO2 threshold is breached in one of the rooms. Then on my work desk (not in a shed, unfortunately) I have a separate ESP32 with a display to provide visual-only alerting there which taps into MQTT to get the metrics (see https://github.com/vladak/workmon).
If the agent could read info like the manual and had a memory system logging vehicle operation data, this seems like it would get pretty useful over time
I wanted something that could run in diverse environments like local LLMs and closed networks, and on self-hosted infrastructure like GitLab and Forgejo.
Works with GitLab, Forgejo, and GitHub Bring your own LLM (Local or external or whatever, any OpenAI compatible endpoint, or Claude API) Reviews pull requests and posts inline comments Replies to users on pull request threads Replies to issues I ran a Martian benchmark with mimo v2.5 pro and placed 7th out of more than 20 code review tools (Precision 0.351, Recall 0.545, F1 0.427). I have not yet tested stronger models. I'll update with results when I do.
Built with Bun, SQLite, and SvelteKit on the dashboard. Everything runs in a single container and deploys in minutes.
I put a lot of time into this and would love for you to try it. Setup takes about 3 minutes only. You can use it for review on your current projects
Feedback, feature requests, and technical questions are all welcome.