Bubo — Code Review Agentico com IA
mountainowl/bubo
Revisão agentica de código para GitHub/GitLab com qualquer LLM
Ver no GitHubServer MCP que implementa revisão de código automatizada e agentica para PRs do GitHub e MRs do GitLab, usando o modelo de linguagem que você escolher (Claude, Codex ou outro). Posta apenas achados acionáveis como threads de comentário inline, sem ruído ou elogios vazios.
Instalação
npx convez-ai mountainowl/bubo1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automatizar code review em fluxos de CI/CD
- Identificar problemas de segurança e qualidade em PRs/MRs antes da merge
- Reduzir carga de revisão manual em equipes de desenvolvimento
- Padronizar feedback técnico com IA no ponto de mudança de código
Veredito

Bubo é uma solução robusta e bem-arquitetada para automação de code review em repositórios. Destaca-se pela abordagem agentica estruturada, suporte a múltiplos LLMs e foco em achados acionáveis — excelente para times que já adotam CI/CD e buscam qualidade sem excesso de ruído.
README
A tradução para o português deste README ainda está na fila.
Bubo 🦉
Agentic AI code review — with the LLM of your choice.
Bubo is an agentic AI code reviewer for GitLab MRs and GitHub PRs. It watches open changes, runs a structured agentic review with the LLM you choose (Codex, Claude, or any model your CLI drives), and posts only actionable findings as inline review threads — no chatbot noise, no praise, no summaries. Like the owl it's named for, it stays silent until it has something worth saying.

📖 Documentation
Full, rendered docs live at → mountainowl.github.io/bubo
(the canonical reference). The docs/*.md files below are the source for that
site; this README is a teaser.
👉 New here? The Recipes (docs/recipes.md) are copy-paste setups for GitLab and GitHub — using Codex (the bundled default) or Claude as the review agent.
What a review looks like
Findings are posted inline in a fixed shape — Issue / Impact / Evidence /
Fix / Confidence:
Issue: HS256 JWT fallback is skipped when Cognito URL construction fails.
Impact: Valid local/shared-secret JWT requests return 500 instead of authenticating.
Evidence: The changed interceptor rethrows InvalidAwsUrlException before fallback runs.
Fix: Treat Cognito validation construction failures as failed Cognito auth when fallback is allowed.
Confidence: 0.94
When a review finds nothing actionable, Bubo posts one short change-level
acknowledgement (Automated review ran — no issues found.) so a clean MR/PR is
distinguishable from one the reviewer never touched. It's default-on, dedup'd by
bot author + exact body, and configurable under [agents] (see the
configuration reference).
Real (sanitized) inline findings on GitLab MRs:


More sanitized examples are in docs/examples/README.md. Demo GIF: docs/media/bubo-demo.gif.
60-second quickstart
Install prereqs (uv, Python 3.14+, Git, plus the CLI for your SCM and a Codex agent — see prerequisites), then:
uv tool install git+https://github.com/mountainowl/bubo@v0.8.0
bubo init # idempotent; --dry-run to preview
# Edit ~/.local/share/bubo/config/env.toml:
# [gitlab].token, [agents].llm_model, [agents].llm_api_key,
# [agents].llm_api_key_env, and at least one [[projects]] entry.
bubo doctor # verify before first poll
bubo-poller # one poll cycle; exits at the end
The first cycle runs with [review].dry_run = true (the default) — findings are
planned but no comments are posted. Flip to false once a real review looks
right. The full walkthrough is in the
Recipes and
install and configure; poller flags and the
bundled MCP server are in run.
Further reading
These render on the docs site and as plain Markdown in the repo:
| Doc | What's in it |
|---|---|
| Prerequisites | macOS / Linux runtime, per-provider tools, credentials, install verification. |
| Install and configure | uv tool install, bubo init, the minimum config/env.toml, GitLab and GitHub bot setup. |
| Run | One-off review, the poller, the bundled bubo-mcp MCP server, and upstream wrappers. |
| Configuration reference | Every [scm] / [gitlab] / [github] / [review] / [poller] / [agents] / [telemetry] / [[projects]] setting and its default. |
| Operate | Remote deploy, scheduling under cron or systemd, --sync-outcomes grading, one-shot backfill. |
| Telemetry | Emitted llm_review.* metrics, ready-made dashboard queries, cardinality discipline. |
Status
- GitLab & GitHub posting via polling — production path, at outcome-metric
parity. Set
[scm].provider = "github"(or runbubo-gh-poller). - MCP server (
bubo-mcp) — read-only metrics + triggered reviews; stdio or HTTP. - Codex or Claude — Bubo runs the review through a wrapper around your agent CLI. Codex ships pre-wired as the bundled default; Claude works the same way once you point the wrapper at it.
- Webhook-driven triggering — not implemented; polling is the only path.
Review execution is intentionally outside CI/CD. Run it as a poller beside your existing pipelines.
Security
config/env.tomlis gitignored and holds tokens. Do not print or commit real values from it.- Review-agent stdout is redacted (
GITLAB_TOKEN=,OPENAI_API_KEY=,glpat-…,sk-…, and credentialed Git URLs) before being written to reports, logs, or the database error column. - The reviewer subprocess is launched under a strict env allowlist — host
secrets are not passed wholesale into the LLM agent. Releases are cosign-signed
with an SBOM. Report vulnerabilities per
SECURITY.md.
Bot avatar
Upload assets/bubo.png as the GitLab (or future GitHub) bot avatar.
![]()