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.

Holt is pre-1.0 and needs Python 3.11 or newer. Pick whichever installer you already have:

pipx install holt-cli          # recommended if you have pipx
uv tool install holt-cli       # if you use uv
pip install --user holt-cli    # plain pip

The same commands work in PowerShell on Windows.

1. Give Holt a GitHub token. Holt only reads public data, so the token needs no permissions. Create one here (leave every box unticked), then:

holt token

It asks for the token and saves it privately on your computer. Already use the GitHub CLI? Holt picks up gh auth token automatically, so you can skip this.

2. Ask about a repository.

holt analyze pallets/flask

That is the whole thing: a verdict and the numbers behind it, free, with no AI key. Or run holt with no arguments for the interactive interface.

export GITHUB_TOKEN=your_token             # macOS / Linux

$env:GITHUB_TOKEN = "your_token"           # Windows PowerShell

With a model set up, the report also explains itself in plain English and quotes the pull request threads it relied on. The model never chooses the verdict. Gemini's free tier is the cheapest way to start:

holt models --provider gemini --model gemini-2.5-flash
export GEMINI_API_KEY=your_key             # PowerShell: $env:GEMINI_API_KEY = "your_key"
holt analyze pallets/flask

Get a Gemini key at https://aistudio.google.com/apikey. Holt also supports OpenRouter (--provider openrouter, key in OPENROUTER_API_KEY), Anthropic, OpenAI, Ollama (local, no key), and any OpenAI-compatible endpoint. Without a model, every command gives the rules-only report.

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

holt

Type a repository and press enter. The first time, it asks for your GitHub token. From there you can revisit past reports, open the evidence behind a claim on GitHub (o), and set up a model (ctrl+l). Press ? on any screen for help.

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.

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

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.

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

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.

Holt is licensed under the Apache License 2.0.