calllint-mcp

The MCP server for CallLint — a static preflight safety gate for MCP servers and agent tools.

Use CallLint before installing or approving other MCP servers. It produces evidence-backed verdicts — SAFE / REVIEW / BLOCK / UNKNOWN — for MCP configs and install snippets, entirely statically. It never executes the server it judges.

calllint-mcp is a thin wrapper: every tool delegates to the same engine behind the calllint CLI. No analysis logic of its own, no network on the scan path, no server execution.

The wrapper is versioned independently of the engine it embeds: calllint-mcp 0.2.0 bundles the CallLint 1.1.x engine. The wrapper version tracks the MCP surface (tools, transport, registry metadata); the embedded engine version tracks verdict semantics.

Install

No install needed — run via npx:

// e.g. Claude Desktop / Cursor MCP config
{
  "mcpServers": {
    "calllint": {
      "command": "npx",
      "args": ["-y", "calllint-mcp"]
    }
  }
}

Tools

| Tool | What it does | | --- | --- | | scan_mcp_config_path | Scan an MCP config file on disk → full ScanReport (verdict + evidence). | | scan_mcp_config_json | Scan MCP config JSON → compact decisions (verdict, fingerprint, reason codes). | | verify_baseline | Compare a config against a recorded baseline → drift / rug-pull signal. | | explain_finding | Full evidence-backed explanation of why a verdict was reached. | | generate_agent_rule | Emit the CallLint agent-safety rule for a host (claude, cursor, copilot, …). | | generate_ci_gate_snippet | Emit a GitHub Actions workflow that gates a repo on its agent-tool surface. |

Guarantees

• Never executes a scanned MCP server (static analysis only).
• No network on the scan path; deterministic given inputs.
• SAFE means "no blockers observed", not proof of runtime safety.
• UNKNOWN is never treated as safe.

Transport

JSON-RPC 2.0 over stdio (MCP). Zero runtime dependencies — the server is bundled into a single file, consistent with the calllint CLI.

License

Apache-2.0