sop-mcp

An MCP server that brings process automation to AI agents through Standard Operating Procedures.

LLMs are powerful but unpredictable when executing multi-step processes — they skip steps, summarize instead of act, and lose track of where they are. sop-mcp solves this by delivering procedures one step at a time, forcing the agent to execute each step and provide concrete output before advancing. This turns SOPs into a control mechanism that makes LLM behavior predictable and auditable.

The result: agents that follow processes the way humans do — step by step, with reasoning enforced at each level.

This approach aligns with Agent SOPs — a standardized markdown format for defining AI agent workflows using RFC 2119 requirement levels (MUST, SHOULD, MAY). sop-mcp adds the execution layer: an MCP server that delivers these procedures one step at a time and enforces completion before advancing.

Install

Or add manually:

{
  "mcpServers": {
    "sop-mcp": {
      "command": "uvx",
      "args": ["sop-mcp"],
      "env": { "SOP_STORAGE_DIR": "/path/to/your/sops" }
    }
  }
}

How It Works

Every session starts the same way — discover what's available, then execute.

list_resources()                                     → catalog of sop:// URIs
run_sop(sop_name="sop_creation_guide")               → Step 1 + instructions
run_sop(..., current_step=1, step_output="...")      → Step 2
run_sop(..., current_step=2, step_output="...")      → Step 3
  ...
run_sop(..., current_step=N, step_output="...")      → Completion

Each response tells the agent to execute the step — not just read it.

Bundled SOPs

Four SOPs ship with the server so new users can try run_sop immediately:

Storage default: ~/.sop_mcp (seeded from the bundled SOPs on first run). Override with SOP_STORAGE_DIR.

Tools

Full parameter reference: docs/mcp-reference.md

Documentation

Development

uv sync              # install dependencies
uv run pytest        # run tests
uv run sop-mcp       # start server locally
uv run python scripts/generate_docs.py  # regenerate docs

License

MIT