Holt

Choose an open-source repository that is worth your time.

Holt reads recent GitHub contribution history and shows what happens when outside contributors try to help: whether they receive a response, whether substantive work gets reviewed and merged, and where newcomer contributions actually land. It turns that evidence into a verdict you can inspect before committing days or weeks to a project.

Holt is read-only. It never opens pull requests, posts comments, or contacts maintainers. Every factual claim in a full report links back to the pull request, review, or comment that supports it.

Winner of Most useful real-world workflow at the micro1 Frontier Engineering Challenge. Holt is now maintained as an independent open-source product by holt-oss.

Install

Holt is pre-1.0. You need uv and Python 3.11 or newer.

uv tool install holt-cli

Create a classic GitHub token with no scopes and expose it to Holt:

export GITHUB_TOKEN=your_token

Start with the rules-only report. It calls no model and needs no model API key:

holt analyze pallets/flask --live --no-model

For a written report with cited findings and quotations, configure a supported model provider and run the full analysis:

export OPENAI_API_KEY=your_key
holt analyze pallets/flask --live

Run holt models to inspect or change the provider. OpenAI, Anthropic, Gemini, Ollama, and OpenAI-compatible endpoints are supported.

Use the terminal interface

Run Holt without a subcommand to open the interactive terminal interface:

holt

From there you can assess a repository, revisit past reports, inspect the evidence behind a claim, compare candidates, and configure model providers.

What you get

A repository assessment answers five practical questions:

• Do first-time contributors get meaningful work merged?
• How quickly does someone usually respond?
• Are pull requests reviewed by people or accepted mechanically?
• Which parts of the codebase have accepted outsider work?
• What specific evidence supports the verdict?

The headline is one of:

Treat the verdict as a filter, not an oracle. Open the cited evidence before making a significant commitment.

Commands

See USAGE.md for the complete workflow and docs/COMMANDS.md for command details.

How it works

Holt separates evidence collection, interpretation, verification, and the final decision:

GitHub history
    ↓
contribution signals + cited findings
    ↓
evidence verification
    ↓
deterministic verdict
    ↓
written explanation

The model may interpret threads and explain the result, but it does not choose the verdict. Unsupported findings are removed before the decision is computed. The same verified inputs therefore produce the same verdict.

The architecture and its boundaries are documented in docs/DESIGN.md.

Evaluation and limitations

Holt has been measured on two temporally held-out repository pools. On the out-of-sample pool it reached 0.82 balanced accuracy and 0.63 Matthews correlation, compared with 0.61 and 0.21 for a single prompt over repository metadata and documentation. The committed fixtures and recordings make those results reproducible.

The evaluation is evidence, not a guarantee. Repository cultures change, quiet periods can look worse than they are, and no benchmark captures whether a particular issue matches your skills. Holt is designed to show its work so you can disagree with it intelligently.

• Evaluation design and limitations
• Reproduce the published results
• Recorded example analyses

Contributing

Holt welcomes focused improvements to evidence providers, verdict rules, terminal UX, documentation, evaluation, and platform support.

git clone https://github.com/holt-oss/holt.git
cd holt
uv sync
uv run pytest -rs

Read CONTRIBUTING.md before opening a pull request. Community participation is governed by the Code of Conduct, and security reports follow SECURITY.md.

License

Holt is licensed under the Apache License 2.0.