Half your model’s context window is burning on tool schemas it will never touch this turn — and you’re paying for every token of it.

The MCP tools tax is quietly killing multi-agent accuracy

Every time an AI agent takes a turn, it sends every connected tool’s JSON schema to the model — even if the task needs exactly one of them. A five-server MCP deployment averages 22,000 tokens of pure tool overhead per turn, and Anthropic’s own engineering data clocks tool definitions at 134,000 tokens before any optimization kicks in.

Three bridge tools replace hundreds of schema lines per turn

Hermes Agent Ships Tool Search for MCP replaces all MCP and plugin tool schemas in the model-visible array with three lightweight bridge tools that fetch only the schemas the model actually needs, on demand, per turn. The model requests a tool by name, the bridge retrieves its schema, and the full definition enters context only at that moment. The result: prompt sizes drop sharply, cache-miss costs fall from $0.07–$0.10 per turn, and Anthropic’s evals record accuracy gains from 49% to 74% on Claude Opus 4.

Agents teams hit the cost ceiling first

• AI engineers running multi-server MCP deployments who are watching per-session costs compound across thousands of daily turns.
• Platform architects who need agent accuracy to hold as tool counts scale past 30 or 50 without rewriting prompt logic.
• Product teams shipping Claude-based agents who lose deals when decision paralysis from irrelevant tool options degrades output quality.

Tool Search is opt-in, which means teams can test it against their existing setup without a full migration.

The context window just became a pricing battleground

With Anthropic, OpenAI, and Google all pushing longer context as a selling point, the real competition has shifted to who can spend that context most efficiently. Agents that waste half their window on unused schemas will lose on both cost and performance as tool libraries grow.

What you can do with it right now

• Connect five or more MCP servers without bloating every prompt turn.
• Cut cache-miss generation costs on high-frequency agent workflows.
• Run Anthropic’s Tool Attention evals against your current deployment to measure overhead.
• Enable progressive disclosure for plugin tools without changing your core agent logic.

Hermes Agent is open-source from Nous Research. Pricing not listed — check our directory.

One real ceiling: Tool Search adds a retrieval step the model must learn to use correctly

If the model misjudges which tool to request by name, the wrong schema loads and the turn fails — an error mode that did not exist when all schemas were visible upfront.

For lighter deployments, LangChain’s tool filtering and OpenAI’s function-calling with manual schema slicing both reduce context overhead. Neither addresses the progressive-disclosure pattern at the MCP protocol level the way Hermes Agent does.

MCP efficiency is becoming the metric that separates production agents from prototypes

The gap between agents that scale economically and those that don’t is narrowing to decisions made at the tool-loading layer right now. We cover tools like this every Friday — subscribe here and we’ll send the best ones straight to you.