Loading ToolBench
High-performance Model Context Protocol (MCP) server connecting Claude Code and AI agents to Google Gemini 3.8 Flash, Gemini 3.1 Pro, reasoning effort, and image generation.
Gemini Relay demonstrates solid tool definition practices with comprehensive schemas and descriptions for most tools. All 12 tools have explicit descriptions and input schemas defined via Zod. However, several tools lack output schema documentation, parameter descriptions are sometimes generic, and error handling guidance is minimal. The gemini-ask tool is particularly complex with 13 parameters, some with interdependencies that are not fully documented. Tools like gemini-ping and gemini-cancel have minimal descriptions (under 100 chars). No tool annotations (readOnlyHint/destructiveHint) are present despite clear risk classifications in the metadata.
Send one prompt to Google Gemini (or the Claude and GPT-OSS models the Antigravity CLI also offers) and get the answer back: questions, code review, analysis, edits, structured JSON. Reference project files with @path to send their contents along. Pass conversationId to continue a thread. For a phased implementation blueprint use gemini-plan; for idea generation use gemini-brainstorm; for everything else use this.
Generate a ranked list of ideas for an open-ended challenge using a named creativity method (SCAMPER, design thinking, lateral, divergent, convergent), with optional feasibility analysis. Ideation only: use gemini-ask for answers and gemini-plan for implementation blueprints.
Stop every Gemini CLI run the relay has in flight right now: a slow gemini-ask, gemini-plan, gemini-brainstorm or gemini-image. Each cancelled call returns an error to its caller. Nothing else is touched.
List recent Gemini conversations the relay can resume: id, last activity, working directory, first prompt. Pass an id to gemini-ask as conversationId to continue that thread. Reads agy's local store; no model turn.
Diagnose and verify Gemini CLI / Antigravity CLI (agy) installation, active backend, CLI version, and system readiness.
Missing output schema documentation. No tool documents what fields/structure it returns. LLMs cannot plan downstream calls or extract required data without knowing response shape.
No tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite clear risk classifications. gemini-ask (WRITE), gemini-image (WRITE), and gemini-cancel (REVERSIBLE) should declare their side effects explicitly.
Parameter interdependencies undocumented. gemini-ask has 13 params with complex relationships (changeMode + chunkIndex + chunkCacheKey; jsonSchema + includeUsage; mode + conversationId). LLMs cannot infer valid combinations.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | <=2025-11-25 | v2 |
Retrieve one chunk of a chunked changeMode reply from gemini-ask, by the cacheKey and chunkIndex that reply reported. No model turn; the cache lives 10 minutes.
Print the active backend CLI's own --help text (agy or gemini), to see which flags the installed version supports.
Generate images from text descriptions through the Antigravity CLI's image generation tool. Supports aspect ratio and output size selection, plus optional export directly into your project.
Lists available Gemini models, default selections, reasoning capabilities, and active backend status.
Liveness check: echoes the prompt back from the relay itself, without touching the CLI or any model. Use gemini-doctor to check the CLI.
Produce a read-only, phased implementation blueprint for one task — steps, dependencies, risks — on Gemini's highest reasoning effort. Never edits files and never answers general questions; use gemini-ask for those.
Test-only: hold a call open for a set number of seconds to exercise progress and timeout handling. Registered only when GEMINI_MCP_TEST_TOOLS is set.
Minimal descriptions on utility tools. gemini-ping (50 chars), gemini-help (55 chars), gemini-doctor (60 chars) lack context for when/why to call them. Descriptions should be 50-200 chars with clear use cases.
No error recovery guidance. Tools do not document what errors are retryable, what requires user intervention, or what alternatives exist. E.g., gemini-ask should guide: 'If model not found, call gemini-models to list available options.'