gitcontribute is available as a CLI, terminal UI, and MCP server for coding agents.

GitHub can show you open issues. It cannot tell you whether an issue is still relevant, already being implemented, appropriate for an outside contributor, or supported by enough evidence to work on safely.

GitContribute collects repository guidance, related issues and pull requests, code context, accepted contribution patterns, and validation results in a local SQLite corpus. Results identify missing or stale coverage.

find work -> understand it -> check competing work -> prove the change -> prepare the handoff

[!IMPORTANT] GitContribute never writes to GitHub. It prepares local research and drafts for you to review.

Quick start

Run the guided setup with Node.js 18 or newer:

npx --yes gitcontribute@latest setup

Choose MCP to use GitContribute from a supported coding agent, CLI for the terminal and TUI, or Both. The wizard shows every planned change before applying it. Adding a repository during setup does not contact GitHub or start a sync.

After setup, start with a repository or an exact issue:

Results include source references, coverage gaps, and suggested next steps.

npm install --global gitcontribute@latest
gitcontribute setup

npm install --save-dev gitcontribute
npm exec -- gitcontribute setup --mode mcp --codex --token-source none --yes

go install github.com/morluto/gitcontribute/cmd/gitcontribute@latest

Contribution workflow

1. Find a candidate

Search a repository or rank its open issues by available evidence, scope, risks, blockers, and signs that maintainer coordination is needed.

Find contribution candidates in golang/go. Exclude issues with active
implementation work and explain why each remaining candidate is worth
investigating.

The ranking only covers stored observations. Results report incomplete coverage.

2. Research an issue

Build a research brief from the issue, repository guidance, discussion, linked work, indexed code, and historical contributions. Extracted maintainer text and checkboxes are not presented as complete acceptance criteria.

Investigate issue owner/repo#42. Summarize the confirmed problem, likely scope,
relevant code, maintainer guidance, and open questions.

3. Check related work

Look for duplicate reports, linked pull requests, closing relationships, and semantically overlapping work before investing in an implementation.

Check whether owner/repo#42 has duplicate reports or competing implementation
work. Tell me what you checked and identify any missing coverage.

Incomplete coverage is reported instead of being treated as proof that no competing work exists. In the MCP workflow, follow the returned typed recovery action, poll the durable job, and reread the affected exact thread or repository before drawing that conclusion.

4. Validate a change

Record a reproduction, test, benchmark, or other validation and compare the unmodified baseline with a candidate. Validation commands run only after explicit approval.

Validate my candidate change against the baseline. Run the approved checks,
record both results, and explain whether the evidence supports the change.

Stored runs include the command, outcome, timing, and available process metrics.

5. Prepare a draft

Create a local issue, pull-request, or review draft tied to the research and validation evidence already collected.

Prepare a pull-request draft for this contribution. Tie its claims to the
recorded evidence, include the validation results, and do not post anything.

Draft revisions retain their exact rendered bytes and provenance.

Interfaces and storage

GitContribute stores repositories, threads, code snapshots, investigations, evidence, validation results, and contribution outcomes in SQLite. Network access is explicit; once information has been synced, corpus search and inspection work offline.

 GitHub read APIs                  Local checkout
       |                                |
       | explicit sync / hydrate        | explicit index / acquire
       v                                v
  +------------------------------------------------+
  |              Local SQLite corpus               |
  | observations · coverage · evidence · outcomes  |
  +------------------------+-----------------------+
                           | offline reads
                 +---------+---------+
                 v                   v
              CLI / TUI          Coding agents

The CLI and MCP server use the same application services and side-effect boundaries.

CLI

The CLI exposes the same workflow without requiring an MCP client:

# Sync repository context and current threads
gitcontribute archive sync-context owner/repo
gitcontribute archive sync owner/repo

# Find and inspect contribution candidates
gitcontribute radar owner/repo --limit 10
gitcontribute research brief issue:owner/repo#42

# Search stored threads and indexed code
gitcontribute search threads "connection timeout" --repo owner/repo
gitcontribute search code "context.WithTimeout" --repo owner/repo

For implementation work:

gitcontribute investigation start-thread issue:owner/repo#42 --json
gitcontribute workspace create <investigation-id>
gitcontribute validation define --kind=test --command="go test ./..." \
  --working-dir=/path/to/workspace <investigation-id>
gitcontribute validation run <validation-id> --kind=base --execute
gitcontribute validation run <validation-id> --kind=candidate --execute
gitcontribute validation compare <base-run-id> <candidate-run-id>
gitcontribute readiness opportunity <opportunity-id>
gitcontribute prepare pr --approach="Bound retries with context" \
  --workspace <workspace-id> <opportunity-id>

Run gitcontribute --help or gitcontribute <command> --help for the complete command and flag reference. Most non-interactive commands accept --json; machine-readable output goes to stdout and progress goes to stderr.

Launch the local TUI with:

gitcontribute tui

MCP

The MCP server advertises one unified catalog. Hosts such as Codex and Claude Code can discover large MCP catalogs with native tool search, so setup does not ask users to choose permanent capability profiles.

gitcontribute setup --mode mcp --codex --token-source none --yes
gitcontribute setup --mode mcp --all-clients --token-source none --yes

To start the stdio server directly:

gitcontribute mcp serve --transport=stdio

Add --read-only to remove tools that permit network access, local writes, or execution. See MCP catalog and composition for capability boundaries, snapshot recovery, resource handoffs, and common tool sequences.

GitHub users are stored as first-class actors. Live discovery records identity only; profile, social-account, organization, pinned-item, repository, and contribution facts are synchronized independently. See the actor corpus for the data and freshness model.

Side-effect boundaries

GitContribute separates corpus reads, GitHub reads, local writes, process execution, and external mutation.

• Corpus reads never fetch data.
• Crawling and indexing never execute repository-controlled code.
• Explicit validation commands run on your host with your user permissions.
• GitContribute has no GitHub mutation capability.
• There is no hosted corpus or product telemetry.

See Architecture for the complete boundary definitions.

Documentation

• Onboarding and configuration
• MCP catalog and composition
• Actor corpus
• Architecture and side-effect boundaries
• Operational runbooks
• Security policy
• Contributing

Storage

GitContribute follows platform conventions:

The default corpus is gitcontribute.db in the data directory. Run gitcontribute metadata or gitcontribute doctor --json to inspect the local setup.

Platform support

Linux and macOS are the primary development and test targets. Windows builds are tested in CI and use the standard %APPDATA% and %LOCALAPPDATA% locations.

Development

make verify
go run ./cmd/gitcontribute --help

Before changing package boundaries or side effects, read docs/architecture.md. See CONTRIBUTING.md for the repository workflow.