code-mode

Why we included this project

Teams building agents that reach many different tools have long dealt with the overhead of function-calling JSON: the model has to ingest hundreds of tool schemas, then make round-trip after round-trip to assemble one result. Code-mode takes a different route by giving the agent a single tool that executes code with access to your registered MCP, HTTP, file, and CLI tools, so a multi-step workflow collapses into one sandboxed code block instead of a string of API calls. That means fewer round trips and less context churn, which is what you want when wiring up real integrations like GitHub PRs or Notion OAuth flows. You can adopt it as a Python or TypeScript library, as a self-configuring CLI for shell-based coding agents, or as a ready-made MCP server for clients such as Claude Desktop, so the entry point can match how your agent already runs. If tool-calling latency or schema bloat is slowing you down, this is a practical thing to test against your current orchestration.