F022: MCP Server Implementation Plan

1. Overview

Implement an MCP (Model Context Protocol) Server for the CSTP Decision Engine. This allows any MCP-compliant agent (Claude Desktop, OpenClaw, other AI assistants) to natively discover and use CSTP capabilities without custom client code.

2. Architecture

• Bridge Pattern: Create a2a/mcp_server.py that wraps the existing Dispatcher.
• Transport: Support Standard Input/Output (stdio) for local agents and SSE (Server-Sent Events) for remote.
• Protocol: Implement MCP v1.0.0-rc.1 spec.

3. Exposed Tools (Resources & Prompts)

The MCP server will expose CSTP methods as Tools:

CSTP Method	MCP Tool Name	Description
cstp.queryDecisions	query_decisions	Find similar past decisions using semantic search
cstp.checkGuardrails	check_action	Validate an intended action against safety policies
cstp.recordDecision	log_decision	Record a final decision to the immutable log
cstp.reviewDecision	review_outcome	Record the outcome of a past decision
cstp.getCalibration	get_stats	Get accuracy and Brier score metrics

4. Implementation Phases

Phase 1: Core MCP Bridge (2 Days)

• [ ] Add mcp python package dependency
• [ ] Create a2a/mcp_server.py using mcp.server.fastmcp
• [ ] Map query_decisions tool to dispatcher.query_decisions
• [ ] Map check_action tool to dispatcher.check_guardrails
• [ ] Verify local stdio connection with Claude Desktop

Phase 2: Full Tool Suite (2 Days)

• [ ] Map log_decision (record)
• [ ] Map review_outcome (review)
• [ ] Map get_stats (calibration)
• [ ] Add input schema validation (Pydantic models)

Phase 3: Docker & Deployment (1 Day)

• [ ] Update Dockerfile to expose MCP entrypoint
• [ ] Add SSE (HTTP) transport support for remote agents
• [ ] Update documentation with "How to connect via MCP"

5. Usage Example (Claude Desktop Config)

{
  "mcpServers": {
    "cstp": {
      "command": "docker",
      "args": ["exec", "-i", "cstp", "uv", "run", "python", "-m", "a2a.mcp_server"]
    }
  }
}

6. Success Criteria

1. Discovery: Agent sees query_decisions and check_action in its tool list.
2. Execution: Agent can successfully query past decisions via the tool.
3. Safety: Guardrails are enforced exactly as they are in the JSON-RPC API.