Feature Request: MCP Server Mode (`catchme mcp`)

[salem221094]

Background

I've been integrating CatchMe with CLI agents (Claude Desktop, Cursor, and Hermes Agent) and the current skill-file approach works — but it's limited to prose answers via catchme ask. Agents can't query specific nodes, get structured data back, or chain multiple calls efficiently without burning extra LLM tokens re-parsing natural language.

The entire agent ecosystem is converging on Model Context Protocol (MCP) as the standard tool interface. CatchMe is architecturally a perfect fit — it's already a local server with structured tree data — but right now it's invisible to MCP-compatible hosts.

The Gap

Current agent integration:

catchme ask -- "What was I working on this morning?"
# Returns: prose string only

Agents receive unstructured text and must re-parse it. There's no way to:

• Get raw structured node data
• Query a specific date range and get JSON back
• Chain list_days → get_session → search_activity in one agent turn
• Expose CatchMe as a tool in Claude Desktop's MCP tool picker

Proposal: catchme mcp Command

Add a catchme mcp command that starts CatchMe as a local MCP server (stdio or HTTP/SSE transport), exposing a small set of tools:

Tool	Description
search_activity(query, date_range?)	Natural language search, returns structured node matches
list_days()	Returns available days with top-level summaries
get_session(session_id)	Returns full session detail with app/location breakdown
get_tree(date)	Returns the full activity tree for a given day as JSON

Example MCP config (claude_desktop_config.json)

{
  "mcpServers": {
    "catchme": {
      "command": "catchme",
      "args": ["mcp"]
    }
  }
}

Once registered, any MCP-compatible agent (Claude Desktop, Cursor, Hermes, etc.) sees CatchMe as a native tool — no skill files, no CLI parsing, no prose round-trips.

Why This Matters

• Zero new capture logic — purely an output/interface layer over existing tree data
• Multiplies CatchMe's reach — every MCP host becomes a potential user automatically
• Structured outputs — agents get JSON they can reason over, not text they must re-parse
• Composable — an agent can call list_days(), then get_session(), then search_activity() in a single turn with full context chaining

Implementation Notes

The MCP Python SDK (mcp[cli]) makes this relatively lightweight — the server just wraps the existing retrieve() generator and tree-loading functions already in catchme/pipelines/. A minimal stdio server could be ~100–150 lines.

Happy to discuss the approach or help prototype it if the direction makes sense.