gemini-direct

The smallest server in the fabric, and deliberately so: one tool, gemini_search, that sends a query to a Gemini model on your own API key and returns the answer together with its token usage. It gives any agent a second model's perspective on demand — useful precisely because it is not the model doing the main work.

Agents reach for it as a cross-check rather than a workhorse: verify a claim before it lands in a report, get an independent summary of a position, or phrase a question outside the current session's framing. The optional system_instruction parameter shapes the second model's role per call, and the returned usage counts keep the cost of every call visible — the platform's stance is that the Gemini tier stays optional and its spend stays observable.

How your agents use it

• "Sanity-check this conclusion before it goes in the report." The agent calls gemini_search with the claim and a focused system_instruction ("act as a skeptical reviewer"), then reconciles the two answers.
• "Get an outside summary of this argument." One call, one independent reading — cheap triangulation for judgment-heavy work.

Prerequisites

A Gemini API key, set once in Settings ▸ API Keys (see the configuration reference). Without the key the tool returns a clean error and the rest of the platform is unaffected — this server is part of the optional Gemini tier, not a dependency. The model is overridable per call.

Tool reference

Tool	Parameters	What it does
gemini_search	query*: string, system_instruction: string, model: string	Query a Gemini model directly. Returns the response text and token usage. Use this for questions requiring a second LLM perspective, documentation verification, or any query that benefits from Gemini's training knowledge.

Where to go next

• BYO Claude & keys — how the platform's bring-your-own-keys model works
• Configuration — where every key lives