DukeDeSouth/tokendiet-mcp

Local MCP server that compresses agent context payloads before they enter the model — with honest BPE counts and safety rollback.

About tokendiet-mcp

TokenDiet is a local MCP server that compresses file reads, command output, search hits, and fetched web pages before they enter the agent context. It uses deterministic transforms (outline, log dedup, snippet caps) and a safety verifier that rolls back when compression would drop protected content or fail to shrink the payload. Token counts use a real BPE encoder (o200kbase by default), not character guesses. Dogfood benchmarks include real M7 agent sessions; internal sprint docs are not published.

This only helps when the agent gets data through TokenDiet tools. If built-in Read or Bash already loaded the full text into context, calling compress afterward cannot undo that cost.

Outline mode returns signatures plus critical comments (SECURITY, TODO, …) when detected. verified: true means structure verified — check warnings and omitted.bodies before editing code. Use expand(ref) or mode=full for audits.

Requires Node.js 20+ and build tools for better-sqlite3 (native addon).

Wire Cursor (example — adjust paths after clone):

Reload MCP servers in Cursor. Point agents at read, run, search, fetch, expand, and stats instead of raw Read/Grep/Bash for large payloads.

Tool Role ------------ read Compressed file read; outline / signatures / symbol for code run Shell command with compressed stdout/stderr search Ripgrep with JS fallback; compressed snippets fetch HTTP fetch with HTML/JSON/text compression expand Full content from a prior ref stats Session and all-time token accounting

From the project README.

Maintaining this server?

Add the radar badge to your README — it shows your project was picked up by MCP Radar and links to this page:

[![On MCP Radar](https://mcp.liqiwa.com/badge.svg)](https://mcp.liqiwa.com/s/DukeDeSouth--tokendiet-mcp.html)

Related MCP servers