@ctok/mcp

MCP server for ctok - lets Claude Code and any MCP-compatible AI agent call ctok as a tool.

Usage

Add to your Claude Code config (~/.claude.json or per-project .claude/settings.json):

{
  "mcpServers": {
    "ctok": {
      "command": "npx",
      "args": ["-y", "@ctok/mcp"]
    }
  }
}

Restart Claude Code. The following tools become available:

| Tool | Description | |------|-------------| | estimate | Token count, cost, model recommendation, and reduction suggestions | | refine | 7-pass prompt refiner - strips filler, improves specificity, saves tokens | | recommend_model | Best Claude model + effort level for a given task | | scan_project | Scan a local directory and estimate its token footprint |

Tool reference

estimate

estimate(prompt, taskType?, pastedCode?, projectContext?)

Returns: { tokens, cost, recommendation, suggestions }

refine

refine(prompt, context?)

Returns: { original, refined, savedTokens, savedPct, specificityScore, passes, warnings }

recommend_model

recommend_model(prompt, taskType?)

Returns: { model, modelReason, alternatives, effort, effortReason, complexity }

scan_project

scan_project(path?, topHeavyCount?)

Returns: { root, projectType, totalFiles, estimatedTokens, byExtension, topHeavyFiles, excluded }

Run directly

npx -y @ctok/mcp

The server speaks JSON-RPC 2.0 over stdio and exits when stdin closes.

License

MIT