Tool Search for Large MCP and Plugin Catalogs

Large MCP and plugin catalogs can spend a meaningful share of every model request describing tools that are irrelevant to the current task. Hermes Tool Search adds progressive disclosure for those deferrable tools.

Official documentation: https://hermes-agent.nousresearch.com/docs/user-guide/features/tool-search

How progressive disclosure works

When Tool Search activates, deferred MCP and non-core plugin schemas are replaced by three bridge operations: search the catalog, describe one selected tool, then call it. Hermes unwraps the call before dispatch, so approval prompts, hooks, guardrails, logs, and activity feeds still see the underlying tool identity.

Core Hermes tools remain directly visible. Tool Search does not hide terminal, file, browser, web, memory, delegation, or the other built-in capabilities.

Configuration checklist

• Leave automatic activation enabled unless measurements justify a change.
• Confirm the active model's context length is accurate.
• Keep search results bounded.
• Use clear MCP and plugin tool names and descriptions.
• Test the search terms users actually request.
• Confirm excluded administrative tools remain unavailable.
• Inspect approval prompts for the underlying tool name.
• Compare latency and token usage before and after activation.

Trade-offs

Tool Search saves repeated schema tokens when many tools are attached and only a few are needed per turn. It adds at least one discovery round trip, so forcing it on for a tiny toolset can be slower and more expensive. Automatic mode activates only after deferrable schemas cross a configured share of the context window.

Pitfalls

• Expecting Tool Search to defer Hermes core tools.
• Enabling it unconditionally for a small catalog.
• Giving tools vague descriptions that cannot be found reliably.
• Assuming a search match means the selected tool is authorized.
• Hiding delete or billing capabilities behind broad search terms.
• Measuring only prompt tokens while ignoring extra round-trip latency.

Verification steps

1. Record the directly loaded tool count and schema size.
2. Attach a sufficiently large test MCP or plugin catalog.
3. Confirm automatic mode activates only above its threshold.
4. Search for one harmless read operation and inspect its description.
5. Call it and confirm logs show the real tool name.
6. Verify hooks, approvals, and post-call handling still run.
7. Compare the same task with Tool Search on, off, and automatic.