mcp-code-execution

Run Serena MCP tools as TypeScript code instead of through the model. Load only the tools you need, process data locally in your script, and return only final results to Claude, cutting token consumption significantly on large datasets and batch operations.

Yes. This Skill lets you import Serena MCP tool wrappers on-demand, execute them as TypeScript, filter and transform data locally, and report only summarized results back. Intermediate processing stays private to your script.

Use this Skill to compose multi-tool workflows—listing files, inspecting symbols, refactoring—entirely in TypeScript code. Process all files locally, aggregate findings, then send only the final summary to the LLM for decision-making.

Place the MCP script in your project context, then execute it via `bun run src/your-script.ts`. The script manages the MCP client lifecycle, loads tools on demand, and runs deterministic steps without requiring model inference on intermediate steps.

Yes. By executing MCP tools as code and processing data locally before reporting results, sensitive intermediate data never reaches the model, making it suitable for cross-project analyses where privacy matters.

This approach works best for deterministic, batch-oriented tasks with clear input-output boundaries. Complex reasoning or dynamic decision-making mid-workflow may still benefit from model interaction; use this Skill when you can define processing logic upfront.