Hermes Agent 推出 MCP 工具搜尋功能：Anthropic 評測顯示 Opus 4 準確率提升 49% 至 74%

Nous Research’s open-source Hermes Agent now ships a Tool Search feature. It directly addresses a growing bottleneck in AI agent systems: too many MCP tools filling up the context window. In this explainer article, we will breaks down what Tool Search does, how it works, and when to use it. The Problem: MCP Tools Are Eating Your Context Window When you connect multiple MCP (Model Context Protocol) servers to an AI agent, every tool’s JSON schema gets sent to the model on every turn. This happens even if the model only needs one or two tools for a given task. Real-world deployments feel this immediately. A Hermes deployment with five MCP servers and 34 tools shows average prompt sizes of 45,000 tokens per turn. Roughly 22,000 of those tokens — around 50% — are tool schema overhead alone. Anthropic’s own engineering data shows tool definitions can consume 134,000 tokens before optimization. Tool Attention measures the “MCP Tools Tax” at 15,000–60,000 tokens per turn for typical multi-server deployments. This creates two distinct problems: Cost: Cache-miss generations at session start can cost $0.07–$0.10 per turn. Accuracy loss: Decision paralysis sets in when the model sees hundreds of irrelevant tool options simultaneously. Source: hermes-agent.nousresearch.com/docs · Nous Research 2026 What is Tool Search? Tool Search is Hermes Agent’s opt-in progressive-disclosure layer for MCP and non-core plugin tools. Instead of loading every tool schema upfront, the model loads only what it needs — on demand, per turn. When Tool Search activates, MCP and plugin tools are replaced in the model-visible tools array by three bridge tools: Copy CodeCopiedUse a different Browsertool_search(query, limit?) — search the deferred-tool catalog tool_describe(name) — load the full schema for one tool tool_call(name, arguments) — invoke a deferred tool A typical interaction looks like this: Copy CodeCopiedUse a different BrowserModel: tool_search("create a github issue") → { matches: [{ name: "mcp_github_create_issue", ... }] } Model: tool_describe("mcp_github_create_issue") → { parameters: { type: "object", properties: { ... } } } Model: tool_call("mcp_github_create_issue", { title: "...", body: "..." }) → { ok: true, issue_number: 42 } The model searches for what it needs, loads the schema, then calls the tool. All hooks, guardrails, and approval prompts run against the real underlying tool name — not against the bridge. The Accuracy Numbers This is not just a token-saving feature. Tool Search also improves model accuracy on MCP evaluations. According to Anthropic’s internal MCP evals: Claude Opus 4: accuracy improved from 49% → 74% with Tool Search enabled Claude Opus 4.5: accuracy improved from 79.5% → 88.1% with Tool Search enabled Large tool catalogs create “decision paralysis” — the model gets confused choosing among many irrelevant options. Removing those options from the context window reduces false positives. Anthropic’s data also shows an 85% reduction in tool-definition token usage while maintaining access to the full tool library. How the Retrieval Works: BM25 + Fallback Under the hood, Hermes uses BM25 — a classic 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 falls back to a literal substring match on the tool name. This protects against zero-IDF degenerate cases, such as searching for "github" in a catalog where every tool name contains “github.” The catalog is stateless across turns. It rebuilds from the current tool-defs list on every assembly. This prevents drift bugs where a stored catalog goes out of sync with the live tool registry. When Does Tool Search Activate? By default, Tool Search runs in auto mode. It activates only when the deferrable tool schemas would consume at least 10% of the active model’s context window. Below that threshold, the tools-array assembly is a pure pass-through. You pay no overhead. This decision is re-evaluated on every turn: A session with just a few MCP tools and a long-context model may never activate Tool Search. A session with many MCP servers attached (15+ tools typically) starts activating it. Removing servers mid-session correctly returns to direct tool exposure on the next assembly. Configuration Reference Add this to your hermes.yaml to control the behavior: Copy CodeCopiedUse a different Browsertools: tool_search: enabled: auto # auto (default), on, or off threshold_pct: 10 # % of context at which auto mode kicks in search_default_limit: 5 max_search_limit: 20 KeyDefaultMeaningenabledautoauto activates above threshold; on always activates if there’s at least one deferrable tool; off disables entirelythreshold_pct10Percentage of context length at which auto kicks in. Range: 0–100search_default_limit5Hits returned when the model calls tool_search without a limitmax_search_limit20Hard upper bound the model can request via limit. Range: 1–50 You can also use a simple boolean shorthand: Copy CodeCopiedUse a different Browsertools: tool_search: true # equivalent to {enabled: auto} Marktechpost’s Visual Explainer #hts-wrap{font-family:'Georgia',serif!important;background:#f7f5f0!important;border:1px solid #e0dbd0!important;border-radius:12px!important;max-width:780px!important;margin:0 auto!important;position:relative!important} #hts-wrap p:empty,#hts-wrap del,#hts-wrap s{display:none!important} #hts-viewport{border-radius:12px 12px 0 0!important;overflow:hidden!important;position:relative!important} #hts-slides{display:flex!important;transition:transform .42s cubic-bezier(.77,0,.175,1)!important;will-change:transform!important;align-items:flex-start!important} .hts-slide{min-width:100%!important;padding:48px 56px 44px!important;background:#f7f5f0!important;position:relative!important;box-sizing:border-box!important} .hts-accent{position:absolute!important;top:0!important;left:0!important;right:0!important;height:3px!important;background:#1a1610!important} .hts-tag{display:inline-block!important;font-family:'Georgia',serif!important;font-size:10px!important;font-weight:400!important;letter-spacing:.12em!important;text-transform:uppercase!important;color:#7a6a52!important;background:#ede8df!important;border:1px solid #d8d0c0!important;border-radius:4px!important;padding:4px 10px!important;margin:16px 0 14px!important} .hts-num{font-family:'Georgia',serif!important;font-size:11px!important;color:#b0a890!important;letter-spacing:.08em!important;margin:0 0 10px!important;display:block!important} .hts-title{font-family:'Georgia',serif!important;font-size:22px!important;font-weight:700!important;color:#1a1610!important;line-height:1.28!important;margin:0 0 16px!important;letter-spacing:-.01em!important} .hts-body{font-family:'Georgia',serif!important;font-size:15px!important;color:#3d3626!important;line-height:1.72!important;margin:0 0 18px!important} .hts-body strong{font-weight:700!important;color:#1a1610!important} .hts-code{display:block!important;background:#1a1610!important;color:#d4c9a8!important;font-family:'Courier New',monospace!important;font-size:12px!important;line-height:1.75!important;padding:18px 22px!important;border-radius:8px!important;margin:0 0 18px!important;overflow-x:auto!important;white-space:pre!important} .hts-stat-row{display:flex!important;gap:12px!important;margin:0 0 10px!important;flex-wrap:wrap!important} .hts-stat{flex:1!important;min-width:140px!important;background:#fff!important;border:1px solid #e0dbd0!important;border-radius:8px!important;padding:16px 14px!important;text-align:center!important;box-sizing:border-box!important} .hts-stat-val{font-family:'Georgia',serif!important;font-size:20px!important;font-weight:700!important;color:#1a1610!important;display:block!important;margin:0 0 5px!important} .hts-stat-label{font-family:'Georgia',serif!important;font-size:11px!important;color:#7a6a52!important;letter-spacing:.03em!important;line-height:1.5!important