OpenCodeReview

AI code review for GitHub pull requests that runs on your own machine, using the Claude, ChatGPT or OpenCode (Go or Zen) plan you already have.

OpenCodeReview watches the repositories you pick, reviews each pull request with Claude Code, Codex or OpenCode inside isolated containers, checks every finding against the code, and posts the result on the pull request.

Features

• Your own AI. Reviews run on the Claude Code CLI (Claude Opus 5.5 at medium effort by default), the Codex CLI, signed in with your Claude or ChatGPT account, or the OpenCode CLI signed in to your OpenCode workspace, on its Go plan or Zen credits. OpenCode's models are listed from its public Go and Zen model lists on opencode.ai, so new ones show up without an OpenCodeReview update. The top bar picks each provider's model and effort. There is no OpenCodeReview account.
• Several accounts with failover. Connect more than one account, of any provider, and pick the one new reviews start on. When an account hits its usage limit, the review continues on the next account of the same provider, then the other providers, and the limited account comes back after its window resets. Settings shows how much of each Claude and ChatGPT plan is left.
• Codebase indexing. Each selected repository is indexed locally: file metadata, a package and import graph, and lightweight symbols. Reviews get the callers, dependencies and tests of the changed code, plus your convention files (AGENTS.md, CLAUDE.md, CONTRIBUTING.md, REVIEW.md and others).
• Usage. See the tokens reviews used by day, pull request, repository, model and account, with findings, time and cache reuse. A pull request opens into each of its runs.
• Review skills. The review methodology is a Markdown skill you can edit, version and roll back in the app.
• Evidence-checked findings. Every finding must quote the exact code it is about and anchor to a changed line. Reports that fail validation are never published as clean.
• Incremental and stacked PRs. New pushes review only what changed since the last review; stacked PRs are reviewed against their immediate base.
• Isolated execution. The AI CLI runs in locked-down runner containers with a read-only snapshot of the code, no repository checkout, and network access limited to your AI provider.
• Guided onboarding at http://localhost:1421: sign in to GitHub, connect your plan, pick repositories and run a first review.

Quick start

git clone https://github.com/BridgLabs/opencodereview && cd opencodereview && docker compose up --build

Open http://localhost:1421 and follow onboarding:

1. GitHub: sign in with the GitHub device flow.
2. Your plan: connect at least one account. Claude uses claude setup-token (you paste back the code it shows); ChatGPT uses the codex login browser flow or a device code; OpenCode signs in to the OpenCode console with a device code (you approve the code it shows), or takes an API key from your OpenCode workspace, which OpenCodeReview checks with OpenCode before saving it.
3. Repositories: choose the repositories to monitor. They are cloned and indexed in the background.
4. First review: pick an open pull request. After that, new pull requests in monitored repositories are reviewed automatically.

The first build compiles the Rust backend and can take several minutes. To skip it, run the images published with each release instead:

docker compose -f docker-compose.yml -f docker-compose.images.yml up -d

Set OCR_VERSION (for example OCR_VERSION=0.1.0) to stay on one release; the default is the newest.

Requirements

• Docker Engine or Docker Desktop with Docker Compose v2 (the Compose file uses additional_contexts), or rootless Podman on Linux as described below.
• A GitHub account with access to the repositories you want reviewed.
• At least one of:
  • a Claude subscription that works with Claude Code (Pro, Max, Team or Enterprise);
  • a ChatGPT plan that includes Codex;
  • an OpenCode workspace with a Go plan (Go models) or Zen credits (Zen models). Go can bill past the plan through Extra Usage; OpenCodeReview can't stop that, so turn Extra Usage off in your Go workspace if you want reviews to stay within the plan.

Local models through Ollama or similar are not supported: the review egress proxy refuses private networks by design.

You do not need Node, pnpm or Rust on the host to run OpenCodeReview. Every AI CLI ships in the image at a pinned, checksum-verified version.

Rootless Podman on Linux

Tested with Podman 5.8.9 and podman-compose 1.6.0, with cgroup v2, Netavark and aardvark-dns. Use these versions or newer; older versions are not validated. The native provider needs version 1.6.0 or newer for service: build contexts. Podman on macOS and Windows is not supported.

Install podman-compose 1.6.0 or newer with your distribution package manager, or run pipx install podman-compose==1.6.0. Select it explicitly: podman compose may otherwise prefer Docker Compose even when podman-compose is installed:

podman-compose --version
export PODMAN_COMPOSE_PROVIDER="$(command -v podman-compose)"
podman compose up --build

Run the complete build, not just build review-runner: the server image must exist before its runner context can be resolved. Keep the provider selected for later podman compose stop, start, restart and down commands. Named volumes preserve data across restarts; down -v deletes that data.

This is the beginning of the README. Read the rest at the original. Read the full README ↗