# Bubo AI Code Review

> Bubo is a self-hosted AI code reviewer for GitHub pull requests and GitLab merge requests. It runs the review model you choose, posts evidence-backed inline findings or an all-clear acknowledgement, and learns from maintainer outcomes to reduce repository-specific noise.

Bubo is an MIT-licensed Python 3.14+ application. Review execution stays on infrastructure and model paths controlled by the operator. The current production trigger is polling; webhook-driven triggering is not shipped.

## Core capabilities

- [Documentation](https://mountainowl.github.io/bubo/overview/): Product overview, security posture, and measured production results.
- [How Bubo works](https://mountainowl.github.io/bubo/how-it-works/): Discover, review, filter, post, persist, and grade maintainer outcomes.
- [Features](https://mountainowl.github.io/bubo/features/): Self-hosting, bring-your-own LLM, confidence gates, verification, governance, OpenTelemetry, and MCP.
- [Recipes](https://mountainowl.github.io/bubo/recipes/): Installation and deployment recipes for GitHub, GitLab, containers, and supported review agents.
- [Configuration](https://mountainowl.github.io/bubo/configuration/): Provider, review, confidence, learning, telemetry, and policy settings.
- [Operations](https://mountainowl.github.io/bubo/operate/): Poller operation, outcome grading, reporting, and governance workflows.
- [Metrics and telemetry](https://mountainowl.github.io/bubo/telemetry/): Review, finding, outcome, latency, token, cost, and privacy controls.
- [MCP server](https://mountainowl.github.io/bubo/mcp/): Read-only metrics and on-demand review tools over stdio or HTTP.

## Maintainer-outcome learning

Bubo fingerprints every finding and later checks how maintainers handled it. Outcomes such as resolved, disputed, deleted, or merged unresolved feed conservative repository-scoped suppression and confidence calibration. This is an auditable control loop around the configured model, not model fine-tuning.

## Installation

```sh
uv tool install bubo
bubo init
bubo doctor
bubo-poller service start
```

The container image is available from `ghcr.io/mountainowl/bubo`. The configured review-agent CLI and its model credentials remain operator supplied.

## Current limitations

- Polling is the production trigger; webhook-driven triggering is not shipped.
- GitHub and GitLab are supported; Bitbucket and Azure DevOps are not supported.
- Pluggable subject-matter specialist Skills are roadmap work, not a shipped capability.
- The public benchmark directory currently provides a reproducible harness with synthetic sample cases; it is not evidence of real-world benchmark performance by itself.

## Primary sources

- [Source repository](https://github.com/mountainowl/bubo): MIT-licensed source, issues, releases, and changelog.
- [PyPI package](https://pypi.org/project/bubo/): Published Python package.
- [Container image](https://github.com/mountainowl/bubo/pkgs/container/bubo): Multi-architecture GHCR releases.
- [Security policy](https://github.com/mountainowl/bubo/blob/main/SECURITY.md): Vulnerability reporting and supported versions.
- [Benchmark harness](https://github.com/mountainowl/bubo/tree/main/benchmarks): Evaluation format, scoring, and caveats.