MCP tools are thin CLI argv wrappers returning wire text inside JSON

[chouswei]

Summary

memnet-mcp does not expose a structured graph API. Each tool builds CLI argv, runs run_memnet, and returns a JSON string whose stdout is still raw wire lines. Agents and domain MCPs re-parse that text.

Evidence

• src/memnet_mcp/server.py: tools call _run([...]) then resp.to_json().
• src/memnet_mcp/client.py: run_memnet(argv) bridges to memnet serve TCP or inline CLI.
• Documented pattern: application-notes/llm-build-on-memnet.md ("Wrap CLI in typed tools; return JSON envelope").
• novel-mcp then parses warm with ad-hoc splitters (see related parser issues).

Impact

Latency and token cost: serialize to TCP to wire text to JSON string to agent parse. No typed fields for LLM tool schemas. Errors live in errors[] and stderr wire, easy to mishandle.

Suggested direction

Optional structured MCP responses (records as objects / TOON) while keeping wire for CLI and snapshots; or in-process store handles for same-process domain MCPs.

[chouswei]

Partial progress (c0bb278): MCP defaults to in-process engine (MEMNET_MCP_TRANSPORT=tcp for serve fallback), cutting TCP round-trips. Tools still wrap CLI argv and return wire text inside a JSON envelope (stdout / stderr / errors) — structured record responses remain outstanding. Leaving open.