Progressive MCP Tool Routing: Stop Drowning Your Agents in 50K Tokens

July 13, 2026 · TormentNexus Team

The MCP ecosystem has over 20,000 registered tools. Dumping all of them into the LLM context window would consume 50,000+ tokens before the agent even starts thinking. Progressive routing solves this.

The Token Budget Problem

Every tool definition costs tokens. A typical tool schema is 200-500 tokens. With 20,000 tools, that is 4-10 million tokens of pure tool definitions — far beyond any model context window.

Progressive Routing: Three Stages

Stage 1 — Semantic Match: Vector embeddings find the top 20 tools most relevant to the current task
Stage 2 — Context Filter: Narrow based on project context, recent tool usage, and agent capabilities
Stage 3 — Lazy Load: Only inject the final 5-10 tool definitions into the context window

The Result

Context window usage drops from 50K+ tokens to under 2K. Agents respond faster, make fewer mistakes, and stay within budget. This is the difference between giving someone a library card and dumping 20,000 books on their desk.

GitHub · Docs