Every token an MCP tool returns is paid for and fills the context window, and real tool output is mostly noise: null fields, hypermedia URLs, repeated blocks and progress-bar spam. This proxy sits between the agent and its MCP servers and shrinks each response by pruning, TOON encoding and local summarization. It talks only to your own MCP servers and a local Ollama.
What it does
Works with any MCP client through a stdio wrapper or a local HTTP proxy: VS Code Copilot, Claude, Cursor, Windsurf, Gemini CLI, Cline and Roo.
Shadow mode measures real savings while still forwarding the originals, so it can roll out with no risk.
Notes name exactly what was pruned, and the agent can fetch the exact original on demand.
A local dashboard shows live traffic, connections, config, eval history and drift warnings, with no CDN.
How it is built
- Lossless first: JSON becomes TOON only after a verified round trip, and duplicate blocks become pointers.
- Fail-open and never bigger: any error, timeout or Ollama outage forwards the original, and an optimized result is used only when it is strictly cheaper.
- Quality gates for CI, a drift watcher and per-client profiles, with hot-reloaded config.
- In the live A/B run, the three files the agent wrote from the optimized result were byte-identical to the originals.
Honest limits
- Token counts are a model-agnostic estimate.
- Numbers come from the repository benchmark, which anyone can re-run.