Hermes Agent Introduces Tool Search to Boost Accuracy and Efficiency

['The open-source Hermes Agent from Nous Research has just gotten a significant upgrade with the introduction of Tool Search, a feature designed to tackle a growing pain point in AI agent systems: the overwhelming number of MCP tools clogging up the context window. This bottleneck has long plagued developers, limiting the efficiency and accuracy of their AI models. Tool Search aims to change that by allowing the model to dynamically load only the tools it needs, on demand, rather than loading every tool schema upfront.', "The problem Tool Search addresses is a pressing one.

When multiple MCP servers are connected to an AI agent, the model receives the JSON schema of every tool on every turn, regardless of whether it needs them. In real-world deployments, this results in substantial token overhead. For instance, a Hermes deployment with five MCP servers and 34 tools experiences average prompt sizes of 45,000 tokens per turn, with approximately 22,000 tokens — or around 50% — dedicated to tool schema overhead alone.

Anthropic's own data reveals that tool definitions can consume as many as 134,000 tokens before optimization, with 'MCP Tools Tax' ranging from 15,000 to 60,000 tokens per turn in typical multi-server deployments.", "Tool Search operates as an opt-in, progressive-disclosure layer for MCP and non-core plugin tools. Instead of loading all tool schemas at once, the model loads only what's necessary on a per-turn basis. When activated, MCP and plugin tools are replaced in the model-visible tools array by three bridge tools.

The model then searches for the required tool, loads its schema, and calls the tool, with all hooks, guardrails, and approval prompts running against the actual underlying tool name. This approach not only saves tokens but also enhances model accuracy on MCP evaluations. According to Anthropic's internal MCP evals, Tool Search yields an accuracy gain of 49% to 74% on Opus 4.", "The benefits of Tool Search extend beyond token conservation.

By reducing the number of irrelevant options in the context window, the feature helps mitigate 'decision paralysis,' leading to fewer false positives. Anthropic's data shows an impressive 85% reduction in tool-definition token usage while maintaining access to the full tool library. Under the hood, Hermes leverages the BM25 information retrieval algorithm to match the model's query against a catalog of tool names, descriptions, and parameter names.

If BM25 returns no positive-score hits, the system defaults to a literal substring match on the tool name, ensuring robust performance even in edge cases.", "Tool Search runs in auto mode by default, activating only when deferrable tool schemas would consume at least 10% of the active model's context window. Below this threshold, the tools-array assembly becomes a pure pass-through, incurring no overhead. Users can control this behavior by adjusting their hermes.yaml configuration file or using a simple boolean shorthand.