MCP Server

The MCP server exposes 19 tools over stdio, enabling Claude and other AI agents to query, annotate, and audit repos conversationally.

Configuration

Claude Code (project-scoped)

Add to .claude.json:

{
  "mcpServers": {
    "repo-knowledge": {
      "command": "rk",
      "args": ["mcp"],
      "env": {}
    }
  }
}

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "repo-knowledge": {
      "command": "node",
      "args": ["node_modules/@mcptoolshop/repo-knowledge/dist/mcp/server.js"]
    }
  }
}

The MCP server reads rk.config.json from the working directory at startup. Ensure the config file exists in the directory where the server process runs.

Knowledge tools

Tool	Description
`get_repo`	Full knowledge dump for a repo
`find_repos`	Filter repos by owner, status, language, framework, shape
`search_repos`	Full-text search across all indexed content
`related_repos`	Show repos related to a given repo
`repos_by_stack`	Find repos using a specific tech stack combination
`repos_needing_work`	Find repos that need attention: stale, unaudited, warnings
`repo_summary`	One-paragraph summary of a repo
`add_repo_note`	Add a typed knowledge note to a repo
`add_relationship`	Record a relationship between two repos
`knowledge_stats`	Database statistics
`sync_repos`	Trigger a full sync cycle
`sync_dogfood`	Sync dogfood evidence from dogfood-lab/testing-os into repo facts

Audit tools

Tool	Description
`audit_submit`	Submit audit results for a repo
`audit_posture`	Get audit posture for one repo
`audit_portfolio`	Portfolio-wide audit posture grouped by health
`audit_findings`	List open findings, filterable by severity and domain
`audit_detail`	Full audit report for a repo
`audit_controls_list`	List canonical controls by domain
`audit_unaudited`	List repos with no audit runs

Multi-agent workflows

repo-knowledge is designed for parallel multi-Claude operations. Multiple agents can:

Audit — claim repos from a worklist, run the 80-control audit, submit via audit_submit
Enrich — add thesis, architecture notes, and relationship mappings
Remediate — fix findings using a scored workflow

See templates/claude-games/ in the repo for full playbook templates.