Takes the insight that LLMs are better at writing code than making tool calls and runs with it. Instead of exposing multiple MCP tools directly, you get one execute_code tool that runs TypeScript/JavaScript in a Deno sandbox. Your code makes HTTP requests to localhost:3001/mcp/* endpoints that proxy to your actual MCP servers. So rather than struggling with tool syntax, Claude writes fetch() calls to chain file operations, database queries, or whatever MCP servers you've configured. The author moved on to a different project called lootbox, but this proves the workflow works for complex tool orchestration.
claude mcp add --transport stdio jx-codes-codemode-mcp uvx codemode-mcp