Designing the MCP server: a cache as a tool the model can hold
MCP turns Crowkis into something an AI assistant can use deliberately, check the cache, store the answer, over plain stdio, with the banner silenced so JSON-RPC stays clean.
The Model Context Protocol reframes integration: instead of your code calling the cache around the model, the model itself holds cache operations as tools. crowkis mcp speaks JSON-RPC over stdio, the MCP transport, and registers lookup, store, and stats surfaces that any MCP-capable assistant can invoke mid-conversation.
Small design decisions carry the production weight. The mcp subcommand boots silently, the startup banner every other verb prints would corrupt a JSON-RPC stream, so it's the one silent door. Tool results are structured for model consumption: a hit returns the answer with its confidence; a miss returns a clean signal the model can act on, not an error to hallucinate around.
- 1agent 1: 'what's the schema?'
- 2crowkis
- 3agent 2: 'schema for orders?'
- 4agent 3: 'show orders schema'
- 5agent 4: 'orders table layout?'
- 6provider
- 7answers
Five agents asking one question should cost one answer.
Crucially, MCP traffic gets no trust shortcuts: an assistant's store request walks the same five-stage pipeline as RESP and SDK writes, with the assistant as a ledger-tracked source. An agent that stores garbage earns a higher bar automatically, the immune system doesn't care who's writing.
The bottom line
The result is agents that remember as a behavior rather than an architecture diagram: Claude Code checks before spending tokens, banks what it computes, and the whole team's assistants share the dividend. One config block; the binary was already running.