Skip to main content

MCP Server

CFA exposes its governance tools via the Model Context Protocol (MCP), allowing any MCP-compatible AI agent (ChatGPT, Claude, Copilot, Cursor, Windsurf) to query and enforce CFA policies in real time.

Quick Start​

pip install cfa-kernel[mcp]
cfa-mcp

Or:

python -m cfa.mcp

Claude Desktop Configuration​

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "cfa": {
      "command": "python",
      "args": ["-m", "cfa.mcp"]
    }
  }
}

Exposed Tools​

1. cfa_evaluate_signature​

Evaluate a StateSignature JSON against the active policy bundle.

{
  "name": "cfa_evaluate_signature",
  "arguments": {
    "signature": {
      "domain": "fiscal",
      "intent": "reconciliation",
      "target_layer": "silver",
      "datasets": [...],
      "constraints": {...}
    }
  }
}

Returns: APPROVE, REPLAN, or BLOCK with faults and remediation steps.

2. cfa_describe_rules​

List all active policy rules with descriptions and severities.

3. cfa_explain_fault​

Explain a fault code — what it means, why it occurs, and how to fix it.

{ "fault_code": "GOVERNANCE_RAW_PII_IN_PROTECTED_LAYER" }

4. cfa_audit_check​

Verify the integrity of the audit trail hash chain.

5. cfa_list_backends​

List registered codegen backends with capabilities.

Protocol​

CFA MCP server uses JSON-RPC 2.0 over stdin/stdout. Zero external dependencies — pure Python standard library.

Client (stdin)  →  JSON-RPC request  →  Server
Server (stdout) →  JSON-RPC response →  Client