Separate tool registration from tool exposure — install broadly, reveal narrowly

Most harnesses conflate “the tool exists” with “the model can see it,” which forces a tradeoff between capability and a bloated, risky tool surface. Hermes splits the two: every tool registers into a central registry at import time, but a separate toolset layer decides what any given run exposes to the model, narrowing further for delegated runs. This lets you keep a broad installed library while holding each run’s model-visible surface small for both token cost and safety — a direct lever against the way Context inefficiency compounds three penalties: cost, latency, and quality degradation, and a complement to how Safety enforcement belongs in tool design, not system prompts. Because Tools are a new kind of software — contracts between deterministic systems and non-deterministic agents, fewer exposed tools also means fewer contracts the model can misuse while working under the The context window is the fundamental constraint — everything else follows.