Your agent's context window is the thing you run out of first. Gemini Relay sends the reading to Google Gemini through the Antigravity CLI, and hands your agent back the answer instead of the files.
Four things a coding agent asks the relay to do, and what comes back.
### 📋 Architectural Blueprint: Resilient WebSocket Server
#### Phase 1: Core Stream Engine & Backpressure
- Implement stream.Writable with highWaterMark: 64KB.
- Buffer overflow strategy: Pause socket frame ingestion when client buffer exceeds 80% threshold.
#### Phase 2: Autonomous Token Rotation
- Background lease manager checking JWT expiry at T-60s.
- Non-blocking handshake refresh without closing established TCP sockets.
#### Phase 3: Failure Modes & Recovery
- Exponential backoff with full jitter: min(30s, 100ms * 2^attempt + rand(0, 50ms)).
📊 [Tokens: 9,240 in, 1,806 out (4,180 thinking)]low, medium and high side by side, and when each one is worth it.
Control thinking token depth per tool call using effort: "low" | "medium" | "high".
Security vulnerability audits, race conditions, distributed architecture design, mathematical algorithms, and comprehensive system planning (gemini-plan).
gemini-plan({
task: "Design distributed lock mechanism with Redis & auto-renew",
context: "High-throughput financial ledger API with strict consistency",
effort: "high"
})What gemini-plan picks when you name no model, and the only model gemini-image uses. ask-gemini and brainstorm send no model at all unless you name one.
Everyday tasks with balanced reasoning depth — bug investigation, refactors across a handful of files, unit tests.
The same Flash model with the least thinking depth baked into the id — for summaries, syntax checks, formatting and commit messages.
Complex problem solving, formal mathematical algorithms, concurrency analysis, and mission-critical architecture.
Pro-grade knowledge retrieval with minimal thinking, for questions that need the Pro model but not its deep reasoning.
Gemini 3.7 and 3.6 Flash are still nameable too, in high, medium and low. The same agy backend also serves claude-sonnet-4-6, claude-opus-4-6-thinking and gpt-oss-120b-medium; pass any of them as model. They draw on a quota bucket separate from the Gemini models. Models lists every id.
Ready? Quick Start has the whole install in three steps.