SDKs & integrations

Crowkis MCP

The Crowkis binary doubles as an MCP server. Claude Code, agent frameworks, and any MCP-capable app can check the cache before calling the model — and store what they compute.

Why this exists#

AI coding assistants and agents repeat themselves: the same documentation lookups, the same explanations, the same boilerplate reasoning, dozens of times a day, billed at full token price every time. With Crowkis behind MCP, the model gains two tools — roughly "check the cache" and "store this answer" — and repeated work becomes a local sub-millisecond hit instead of a paid round-trip.

In plain words: your AI assistant gets a memory that lives on your machine. Questions it has already answered stop costing you money. Nothing is sent anywhere except to the model you were already using — and only for genuinely new questions.

Set it up#

crowkis mcp speaks JSON-RPC over stdio — the standard MCP transport. It boots silently (no banner) so the protocol stream stays clean. Register it in your client's MCP configuration:

claude code — .mcp.json

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

Or add it from the command line:

shell

claude mcp add crowkis -- crowkis mcp

Running Crowkis in Docker? Point MCP at the container instead:

docker variant

{
  "mcpServers": {
    "crowkis": {
      "command": "docker",
      "args": ["exec", "-i", "crowkis", "crowkis", "mcp"]
    }
  }
}

What the model sees#

Tool surface	What it does
Cache lookup	Semantic CGET against the running server — paraphrases of previously answered questions hit, with the same confidence and safety gates as every other client.
Cache store	CSET with model and tenant attribution, so agent answers pass the same anti-poisoning pipeline before they're trusted.
Stats	Hit rate and savings, so the agent (and you) can see what the cache is worth.

Every MCP-originated entry goes through the identical five-stage trust pipeline as RESP or SDK traffic. An agent can't poison the cache any more easily than a human can.

Patterns that pay off#

Workload	Why it caches well
Doc & API lookups	The same 'how does X work' questions recur across a team all day.
Code explanation	Explanations of stable code are stable — cache until the file changes, then invalidate.
Agent tool results	Deterministic tool calls (search, schema fetch) are pure savings on replay.
Multi-agent fan-out	Five agents asking variants of one question become one model call and four hits.