muxed

> Aggregate all your Model Context Protocol (MCP) servers behind a single daemon. Fewer tokens. Faster execution. Better accuracy.

muxed is a background daemon and CLI that sits between your AI agent and your MCP servers. It solves the problems that Anthropic and Cloudflare have been writing about: tool sprawl eating your context window, slow cold starts, and degraded accuracy as you add more servers.

The MCP ecosystem has a scaling problem. Every tool you connect dumps its full schema into the model's context window. A standard setup with 3-4 MCP servers can consume 20-30% of the context before the agent even starts working. Anthropic's research shows this leads to a 98.7% token overhead on intermediate results. Cloudflare found that agents can't reliably handle more than a handful of servers before tool selection accuracy collapses.

More tools = worse results. Every token spent on MCP tool schemas is a token not spent on your actual task – or on the skills, prompts, and default tools that agents execute deterministically.

muxed is an optimization layer for your MCP infrastructure:

• Fewer tokens in context – Tools stay in the daemon, not in the prompt. Agents discover tools on-demand with muxed grep and muxed info instead of loading every schema upfront. Load only what you need, when you need it.
• Faster execution – Servers stay warm in a background daemon. No cold starts, no repeated connection negotiation. Call tools directly via CLI without round-tripping through the model.
• More precise tool selection – By offloading tool management to muxed, your agent's context window stays clean for what actually matters: reasoning, prompts, and the task at hand. Fewer tools in context means the model picks the ri