Agent Readiness

@alipajand/agent-readiness-mcp

A local Model Context Protocol server that audits repositories for AI coding agents with two deterministic engines:

• agent-readiness-kit: a 0–100 readiness score across 13 categories (instructions, architecture, workflow, testing, safety, and more) for Node.js, Python, Go, and Rust projects.
• agent-context-doctor: quality and safety issues in AGENTS.md, CLAUDE.md, Cursor and Copilot rules, and agent MCP/settings files. Credentials quoted in any finding are redacted.

The audit is read-only and offline. It uses no LLM and no API key, sends no telemetry, and never runs the audited repository's scripts. Both engines are compiled into the package from pinned commits, so it has no runtime dependencies.

Tools

audit_repository and audit_agent_context take an optional path (default: the first root). All three tools are annotated read-only.

Install

Requires Node.js 22.13 or later (23.5 or later on Node.js 23). Older releases have no stable permission model for the audit sandbox described below, and the server refuses to start on them.

npm install -g @alipajand/[email protected]

Configure a client

The server only audits directories inside its roots. Give one --root per directory the client may audit; with none, it uses the directory it was started in.

Claude Code:

claude mcp add agent-readiness -- agent-readiness-mcp --root /path/to/projects

Cursor (.cursor/mcp.json) and other clients that use the mcpServers format:

{  "mcpServers": {    "agent-readiness": {      "command": "agent-readiness-mcp",      "args": ["--root", "/path/to/projects"]    }  }}

Without a global install, use "command": "npx" with "args": ["-y", "@alipajand/[email protected]", "--root", "/path/to/projects"]. Keep the version pinned.

Options: --root <dir> (repeatable), --timeout <seconds> (default 120), --max-concurrency <n> (default 2).

Security

• Paths are resolved through symlinks and must stay inside a root.
• Each audit runs in a separate Node.js process started without a shell, with an empty environment, a time limit, and an output limit. That process runs under Node's permission model: it can read only the audited directory and its own script, and cannot write files or start processes. Node lets a granted read follow a symlink, so the engines' own symlink checks still contain reads; the sandbox is a second layer.
• Stdout carries only MCP messages; diagnostics go to stderr.
• Reports are kept in memory for the session (the 20 most recent) and never written to disk.
• Text in results comes from the audited repository and should be treated as data.
• The server is offline, but the MCP host is not part of it: the host's model reads the results and the host may send them, or files it reads itself, to its model provider. Credential masking covers the server's results only.

Listings

Listed on MCPRush.

License

MIT. Bundled third-party licenses are in dist/THIRD-PARTY-LICENSES.txt and dist/audit-worker.THIRD-PARTY-LICENSES.txt.