PR-Agent

PR-Agent (Qodo) reviews pull requests with an LLM. Its commands are /review, /describe, /improve and /ask. It runs as a GitHub Action, a webhook server or a CLI.

Upstream: source, docs.

Plugboard options

Give PR-Agent its own key. The review sends the whole diff and asks for a YAML reply, so use a model with a long context. The runner must reach plugboard (Serving other machines).

{
  "options": {
    "listen": { "host": "0.0.0.0", "port": 4010 },
    "keys": { "pr-agent": "{env:PLUGBOARD_KEY_PR_AGENT}" },
    "models": ["anthropic/claude-sonnet-4-6"],
    "aliases": { "review": "anthropic/claude-sonnet-4-6" },
    "apps": { "pr-agent": { "allowedModels": ["review"] } }
  }
}

Connect

Put the key in .secrets.toml (or in GitHub secrets for the Action):

[openai]
api_base = "http://192.0.2.10:4010/v1"
key = "<the pr-agent key>"

Or in the environment:

# Double underscore separates the section from the key (upstream convention).
OPENAI__API_BASE=http://192.0.2.10:4010/v1
OPENAI__KEY=<the pr-agent app key>

Use the address your runner reaches in place of 192.0.2.10. Set the model in configuration.toml:

[config]
model = "openai/review"
fallback_models = ["openai/review"]
custom_model_max_tokens = 200000  # the context window of the model behind the alias

The openai/ prefix routes through LiteLLM to an OpenAI-compatible URL, and plugboard receives the model name without it (review). PR-Agent does not know a prefixed model's context size. Without custom_model_max_tokens the review stops before it sends a request, with Ensure openai/review is defined in MAX_TOKENS ... or set a positive value for it in config.custom_model_max_tokens.

Review a local branch

PR-Agent's local git provider needs no hosted pull request. It compares the checked-out branch with a target branch in the current repository and writes the review to review.md. It takes the target branch where the other providers take a pull request URL. The same settings as above, as environment variables:

env = {
    **os.environ,
    "CONFIG__GIT_PROVIDER": "local",
    "OPENAI__API_BASE": os.environ["PLUGBOARD_BASE_URL"],
    "OPENAI__KEY": os.environ["PLUGBOARD_KEY"],
    "CONFIG__MODEL": f"openai/{model}",
    "CONFIG__FALLBACK_MODELS": json.dumps([f"openai/{model}"]),
    "CONFIG__CUSTOM_MODEL_MAX_TOKENS": "200000",
}
# Run from inside the repo. The local provider takes the target branch where the other providers take a pull request URL.
run = subprocess.run([sys.executable, "-m", "pr_agent.cli", "--pr_url", "main", "review"], cwd=repo, env=env, capture_output=True, text=True)

A review that fails still writes review.md, with the text Failed to review PR. Read the file for the model's findings.

A runnable version that builds a throwaway repository with a main and a feature branch is in examples/pr-agent/. It needs no GitHub token.

What works

/review, /describe, /improve and /ask, each as a chat request through config.model, on a local branch or on GitHub, GitLab, Bitbucket and the other hosted providers, which take the same model settings.

Limits

• Keep the key in .secrets.toml or a secret store, not in the repo.
• A model that is not from OpenAI may need extra parameters. Run one /review before you rely on it.