Use Clarx inside Cursor, Claude Code, and any MCP-capable coding agent — score manifests, validate clarx-manifest.json, and generate CI workflows without leaving the IDE.

Clarx MCP Server

@clarxai/mcp puts Clarx's manifest tooling inside your coding agent via the Model Context Protocol. Your agent can score CLAUDE.md / AGENTS.md files, validate clarx-manifest.json, generate CI workflows, and (with your Anthropic key) write manifest sections — all as native tool calls.

No account required. The core tools run entirely on your machine. Files are read locally and never uploaded.

The server is listed in the official MCP registry as ai.clarx/mcp.

Install

Cursor

Add to .cursor/mcp.json in your repo (or ~/.cursor/mcp.json for all projects):

{
  "mcpServers": {
    "clarx": {
      "command": "npx",
      "args": ["-y", "@clarxai/mcp@latest"]
    }
  }
}

Claude Code

claude mcp add clarx -- npx -y @clarxai/mcp@latest

Claude Desktop

Add the same JSON block as Cursor to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows).

Enable the AI writing tools (optional)

Two additional tools — suggest_manifest_fix and generate_manifest_draft — call the Anthropic API with your own key. Add it to the server environment and they register automatically:

{
  "mcpServers": {
    "clarx": {
      "command": "npx",
      "args": ["-y", "@clarxai/mcp@latest"],
      "env": { "ANTHROPIC_API_KEY": "sk-ant-…" }
    }
  }
}

Without the key, the AI tools don't register at all — they never occupy your agent's context, and nothing breaks.

The manifest loop

Talk to your agent in plain language; it picks the right tool:

No manifest yet? — "Draft an AGENTS.md for this repo from the README". The draft is grounded in your README; anything the model can't ground is marked with a TODO comment rather than invented.
Score it — "Score my CLAUDE.md with clarx". Returns a 0–100 quality estimate, five pillar scores, and findings with line numbers.
Close findings — "Write the missing Testing section". Produces one validator-passing markdown section to insert.
Declare intent for the engine — "Create a clarx-manifest.json and validate it". Checks the keys the engine reads: generated, highFanIn, highFanOut, verificationCommands, and more. Committing this file closes O1 and makes the Boundary Clarity pillar evaluable.
Wire up CI — "Add a Clarx PR gate at min score 70". Generates the GitHub Actions workflow — see CI Integration.

Connected mode (Clarx Cloud)

With a Clarx Cloud account, create an API token (org settings → API tokens) and add it to the server environment:

      "env": { "CLARX_MCP_TOKEN": "clarx_…" }

Four hosted tools register — list_repos, get_scan, list_findings, and get_remediation_prompt — so your agent can pull real scan results and the dashboard's remediation prompts without leaving the IDE: "What did the last clarx scan find?", "Get the remediation prompt and fix the top finding." See the tools reference for details.

Two different scores

The MCP analyze_manifest tool returns a manifest quality estimate — fast heuristics over a single markdown file. The CLI returns the repo AI-readiness score — the full engine over your whole codebase. They use different rule systems and are not comparable: a 90 manifest can live in a 60 repo. Use the estimate to write good manifests; use the CLI score as your quality gate.

Every analyze_manifest result carries score_type: "manifest_quality_estimate" so agents don't conflate the two.

Next steps

Tools reference — inputs and outputs for all five tools
CI Integration — the PR gate the MCP server can generate for you
The manifest — what a great agent manifest contains

MCP Server

On this page