Hermes Tool Router

What Hermes Tool Router does

Hermes Tool Router reduces first-turn tool-schema overhead without changing the tool prefix repeatedly later in the conversation. It classifies only the first turn. High-confidence routes send a narrower set of tools to the main model, and uncertainty keeps the full surface. If a pruned tool is needed later, Hermes adds its owning toolset for the rest of the session, through a request_toolset tool or automatic recovery in the tool_request middleware.

Routing can use deterministic rules, an optional OpenRouter classifier, or both. The README gives a config example with a 0.90 confidence threshold, a 2000 ms hard timeout and google/gemini-3.5-flash-lite as the classifier model. It fails open on a timeout, malformed output or low confidence. The implementation is plugin-only, needs no Hermes core patch, and python diagnostics.py reports whether a routing path exists before the provider request.

Key features

• First-turn tool routing that narrows the schema sent to the main model
• Session-sticky tool surface with permanent, monotonic recovery of requested toolsets
• Fail-open behavior on uncertainty, timeouts, invalid classifier output or unsupported runtimes
• Optional OpenRouter classifier, disabled by default
• request_toolset fallback tool generated from the live toolset registry
• diagnostics.py that exits 0 when a routing path is available and 2 when it is not

When to use it

• Cut input tokens and first-response latency when a large tool catalog is loaded
• Test model-driven tool routing on a separate Hermes profile before wider use
• Check whether your Hermes runtime supports first-turn tool pruning

Who it is for: Hermes Agent users with large tool catalogs who want lower first-turn token cost and are willing to test an experimental plugin.

How it fits with Hermes Agent

It is a standalone Hermes Agent plugin that works through stock Hermes hooks such as pre_llm_call and the tool_request middleware, with no change to Hermes core.

Requirements: A Hermes Agent runtime with the needed hooks; an OPENROUTER_API_KEY only if you enable the optional classifier

Note: The author labels it experimental and advises testing on a separate Hermes profile first.

FAQ