Using RCG with your coding assistant (MCP)¶

RCG ships an MCP server, rcg-mcp, so your AI coding assistant can check your agent rule corpus for contradictions before it acts — and answer "which rules fire for this action, and do they conflict?" on demand.

This guide gives copy-paste setup for the popular tools.

What you get — three tools¶

Tool	What it does
`check_corpus(path, provider="mock", semantic=False)`	Run the conflict passes over a rules directory → findings + coherence score.
`explain_action(action, path, scope="*", provider="mock")`	"Which rules fire for this action, and do they conflict?" + a verdict.
`score_corpus(path, provider="mock")`	Just the coherence score + by-type breakdown.

By default the server uses the offline heuristic extractor — no API key, and nothing leaves your machine. For LLM-quality extraction/judging, set ANTHROPIC_API_KEY in the server's environment and pass provider="anthropic".

Step 1 — install the server (once)¶

Recommended (stable, pinned):

pipx install 'rule-coherence-graph[mcp]'     # provides the `rcg-mcp` command

Zero-install alternative (needs uv; re-resolves each launch):

uvx --from 'rule-coherence-graph[mcp]' rcg-mcp

PATH gotcha: GUI apps (e.g. Claude Desktop) often don't inherit your shell PATH, so they may not find rcg-mcp in ~/.local/bin. If a client says "command not found," use the absolute path from which rcg-mcp (or which uvx), or use the uvx form below.

The universal config¶

Most clients accept this shape (stdio transport):

{
  "mcpServers": {
    "rcg": { "command": "rcg-mcp" }
  }
}

Zero-install variant:

{
  "mcpServers": {
    "rcg": {
      "command": "uvx",
      "args": ["--from", "rule-coherence-graph[mcp]", "rcg-mcp"]
    }
  }
}

To enable LLM-backed extraction, add an env block to the server entry:

"env": { "ANTHROPIC_API_KEY": "sk-ant-..." }

Per-assistant setup¶

Claude Code (CLI)¶

# zero-install
claude mcp add rcg -- uvx --from "rule-coherence-graph[mcp]" rcg-mcp
# …or, if installed via pipx
claude mcp add rcg -- rcg-mcp

- Scope: add -s user to enable it in every project, or -s project to share it via a committed .mcp.json. - Verify with claude mcp list, or run /mcp inside a session. - Then ask: "Use the rcg tools to check .agent/rules for conflicts."

Claude Desktop¶

Edit the config file: - macOS: ~/Library/Application Support/Claude/claude_desktop_config.json - Windows: %APPDATA%\Claude\claude_desktop_config.json

Add the universal mcpServers block. Use the uvx form or an absolute path to rcg-mcp (Desktop doesn't see your shell PATH). Restart Claude Desktop — the tools appear under the tools/🔌 menu.

Cursor¶

Create .cursor/mcp.json (project-scoped) or ~/.cursor/mcp.json (global):

{ "mcpServers": { "rcg": { "command": "uvx", "args": ["--from", "rule-coherence-graph[mcp]", "rcg-mcp"] } } }

Open Settings → MCP — "rcg" should appear; toggle it on (green dot). In Agent/Composer, ask it to run the rcg tools.

VS Code (GitHub Copilot — Agent mode)¶

VS Code uses the servers key (not mcpServers). Create .vscode/mcp.json:

{ "servers": { "rcg": { "type": "stdio", "command": "uvx", "args": ["--from", "rule-coherence-graph[mcp]", "rcg-mcp"] } } }

Or run Command Palette → "MCP: Add Server". Requires Copilot Chat in Agent mode; the tools then show in the tools picker.

Windsurf (Codeium)¶

Edit ~/.codeium/windsurf/mcp_config.json (or Cascade → MCP → Manage/Configure):

{ "mcpServers": { "rcg": { "command": "uvx", "args": ["--from", "rule-coherence-graph[mcp]", "rcg-mcp"] } } }

Click Refresh in the Cascade MCP panel.

Cline (VS Code extension)¶

In the Cline panel → MCP Servers → Configure MCP Servers (opens cline_mcp_settings.json) → add the universal mcpServers block. "rcg" and its tools appear in the list.

Zed¶

Zed uses context_servers in settings.json:

{ "context_servers": { "rcg": { "command": { "path": "uvx", "args": ["--from", "rule-coherence-graph[mcp]", "rcg-mcp"] } } } }

Verify it works¶

Ask your assistant something like: - "Use the rcg check_corpus tool on .agent/rules and summarize the conflicts." - "With rcg explain_action, what rules fire if I deploy to production, and do they conflict?"

You should see the tool get invoked and a findings/score result come back.

Notes & troubleshooting¶

Offline by default. provider="mock" needs no key and sends nothing externally. Pass provider="anthropic" (+ ANTHROPIC_API_KEY in the server env) for LLM-quality results.
"command not found". GUI clients may not see ~/.local/bin — use the absolute path (which rcg-mcp) or the uvx form.
Point it at your rules. Pass the path to your corpus (e.g. .agent/rules, .cursor/rules, CLAUDE.md, or . to scan the repo).
Config formats evolve. If a client renamed a key, check its current MCP docs — the command/args for rcg stay the same.

Want conflict checks in CI instead of (or alongside) your editor? Use the reusable GitHub Action — see the README's "Use it in CI" section.