mcp-fusion: Route 5,000+ AI tools into one. Stop LLM context bloat.

Hey Product Hunt! 👋 I’m Renato from Vinkius Labs, and today we’re completely open-sourcing our core routing engine: mcp-fusion. We’ve been building massive Agentic workflows using the new Model Context Protocol (MCP) with Claude and Cursor. But we hit a brutal architectural wall: as soon as we exposed our internal CRUD APIs (50+ endpoints) as individual MCP tools, the entire system collapsed. 1️⃣ API costs exploded: Sending 50 JSON schemas on every single LLM prompt burned our token budget just to remind the AI what tools existed. 2️⃣ Routing Hallucinations: The LLM got confused by semantic overlap, often picking 'update_project' when it should pick 'create_project', and inventing parameters. We couldn't find a solution, so we built mcp-fusion. It acts as an enterprise-grade build-time multiplexer. Instead of feeding the LLM 50 fragile tools, mcp-fusion bundles them into ONE single tool behind a strict discriminator field. The "black magic" under the hood: ⚡️ Zero-Prompting: It uses Zod AST introspection to auto-generate mathematically perfect routing hints for the LLM. 🛡️ Security Boundary: Strict Zod stripping ensures hallucinated parameters never reach your handlers. 📉 Cost Reduction: It compresses metadata using TOON formatting to slash your token footprint even further. If you are building AI Agents or scaling Cursor in your team, this will save you a massive amount of Anthropic/OpenAI API costs and eliminate routing errors. It’s 100% Open-Source.