| Hey HN! I built pctx, a framework that lets AI agents execute code directly instead of making tool calls to save on tokens on MCP. I've been working on code generation in Rust for 2 years, mostly fighting with OpenAPI specs that are either missing or incorrect. MCP's self-documenting nature made this a fun project to tackle since the servers describe their own capabilities accurately. I could focus on building execution sandboxes instead of wrestling with bad specs. After seeing how Cloudflare and Anthropic's Code Mode implementations reduced tokens, I decided to build an open-source version with a focus on correctness and local-first design. pctx prioritizes reliability (TypeScript validation pre execution) and DevEx (no dependencies, 100% local, built-in MCP auth) Technical details:
- Two locked-down Deno sandboxes:
- TypeScript compilation/validation (pre-execution)
- Execution with MCP network access only
- Compiles to a single binary, no dependencies
- Built-in TypeScript code generator and MCP auth utilities
- Uses the official rmcp client/server within the execution runtime Coming soon:
- Automatic authentication of the pctx server
- Python and TypeScript SDKs for using natively in agent frameworks
- Anything else people want
- One-click cloud deployment (see website) GitHub: https://github.com/portofcontext/pctx
Website: https://portofcontext.com Would love your feedback on this. Thanks! |