Contribos

What the contribos MCP server does

The contribos MCP server helps contributors inspect a repository before changing code and prepare evidence for a maintainable contribution. Its tools cover repository policy, issue discovery, contribution proposals, code-reading briefs, local setup, failure diagnosis, pre-submit checks, review responses, and tone checks.

It is deliberately advisory. Contribos does not open pull requests, post comments, claim issues, or write the contributor's explanation. Instead, it surfaces repository evidence and produces drafts or checks that the contributor can inspect and decide whether to use.

How it works

The server runs as a dependency-free MCP process over stdio. Repository data is gathered from cached git clones, file contents, contribution documents, CI configuration, templates, and project history. A SQLite index records main-line changes, affected files, and linked issues so related changes and tests can be identified.

The policy tool quotes repository rules with file and line references where available. Brief generation ranks likely files using terms from the issue, paths, and similar historical changes, then looks for related tests. The check workflow compares the proposed scope with past changes, verifies tests, checks sign-off and disclosure details, and can run a test both with the change and with source files restored to the base version.

Review analysis classifies reviewer comments, relates them to code and repository rules, and tracks unanswered items. Optional language-model summaries use only evidence already collected and can be configured for Anthropic or OpenAI.

Setup and configuration

Install the Python package with pip install contribos. The package requires Python 3.10 or newer and git, and the README reports zero external dependencies. Start the MCP server with:

python -m contribos mcp

For Claude Code, the documented registration form is claude mcp add contribos -- python -m contribos mcp. The server caches clones and indexes under ~/.cache/contribos; set CONTRIBOS_CACHE to change that location.

Set GITHUB_TOKEN for issue- and pull-request-related operations such as find, claim, brief from an issue URL, review, and record. Policy inspection can run without a token, although pull-request signals are omitted. Offline-capable workflows include policy, title-based brief and claim operations, setup, check, benchmark, and review with supplied data.

Optional summaries use CONTRIBOS_LLM=anthropic with ANTHROPIC_API_KEY, or CONTRIBOS_LLM=openai with OPENAI_API_KEY. CONTRIBOS_MODEL can select the model.

Tools and capabilities

The contribos MCP server exposes nine tools:

• policy examines AI policies, disclosure formats, contribution gates, tests, changelogs, DCO or CLA requirements, style tools, and activity.
• find ranks open issues and reports repository welcome signals and responsiveness.
• claim drafts a pre-coding proposal and, when relevant, a vouch-list introduction.
• brief identifies files, similar changes, tests, review history, and likely questions.
• setup and diagnose derive setup instructions and categorize failures.
• check performs pre-submit checks and can draft a pull-request description.
• review analyzes review comments, reply needs, and CI failures.
• tone checks prospective text for machine-written signals, length, missing questions, and placeholders.

Limitations and notes

Results depend on the repository’s available documentation, history, CI configuration, and GitHub access. The README’s benchmark uses historical pull requests as issue stand-ins, so it does not establish performance on real issue text and review comments. Contribos provides evidence and drafts; contributors remain responsible for decisions and for any publication or pull request activity.