Show HN: MCP-compress-router – MCP Compressor
MCP-compress-router is a tool that reduces token usage when using multiple MCP servers with an LLM. Instead of exposing all tool descriptions on every request, it proxies servers through just two tools—get_tool_schema and invoke_tool—cutting overhead from ~26K tokens to ~900–2,000 tokens, saving over 90% on token costs.
Background
- **MCP (Model Context Protocol)** is an open standard that lets AI coding assistants (like Claude, VS Code's Copilot, etc.) connect to external tools — e.g., a GitHub client, a Notion reader, or a code analyzer (Pylance). Each tool must be described to the LLM in full (name, what it does, its parameters) so the model knows when and how to call it.
- **Token limits and cost**: LLMs charge per token (roughly per word fragment). Every time you chat with the model, the entire conversation history plus all tool descriptions must be re-sent. With multiple MCP servers, those tool descriptions pile up fast — 26,000 tokens (~$0.93 per session at Anthropic's Opus pricing) just for overhead.
- **mcp-compress-router** solves this by acting as a middleman: instead of dumping every tool's full schema into every request, it exposes only two generic tools. The LLM first asks for a specific tool's schema only when it actually needs it, then invokes it. This cuts the per-request overhead by 90%+ without losing functionality.
- The project was posted as "Show HN" (a Hacker News tag for sharing your own work), and the author is the same person behind the popular **AdGuard** DNS ad-blocker.