Agentic query: Gemini runs inside your workingDirectory with read_file, grep, list_directory, and glob tools. Pass file paths as hints (not content) — Gemini reads them itself and can explore surrounding code for context.

Capabilities:

• Code analysis with full repo exploration (Gemini follows imports, reads tests, checks related files)

• Image understanding: screenshots, diagrams, architecture charts (png/jpg/gif/webp/bmp)

• General knowledge questions and technical research

• Text transformation, summarization, and generation

• Change mode: structured edit blocks parsed into a machine-applicable edits array (see 'changeMode' below)

File handling: Pass file paths in the 'files' array as hints. Text files are referenced via @{path} — Gemini reads them with its own tools. Image files use --yolo mode for native pixel access. Gemini may also read files beyond the ones you hint at.

Note: Gitignored files cannot be read in text-query mode (plan mode restriction). Image queries (--yolo) can read gitignored files.

Change mode: set 'changeMode: true' to ask Gemini to emit structured **FILE: <path>:<start>-<end>** / ===OLD=== / ===NEW=== blocks instead of prose. The response text stays in response; parsed edits are returned on _meta.edits and never chunked. The tool runs in default agentic mode (NOT plan mode, which refuses to emit edit blocks) with a pre/post-spawn git snapshot that detects any file writes Gemini might attempt. If writes are detected the tool returns _meta.appliedWrites: true and omits edits for safety. Text-only (image files rejected). Requires a git working directory.

Model tips: Use gemini-2.5-flash for speed, gemini-2.5-pro for depth and complex reasoning. If omitted, the CLI auto-selects via its routing model.

Each invocation spawns a fresh CLI process (~15-20s startup overhead). Plan timeouts accordingly.

Google Search grounded research. Gemini searches the web using google_web_search and synthesizes a comprehensive answer with source URLs and citations.

Use for: current events, documentation lookups, API references, comparing technologies, verifying facts, finding recent releases or changelogs.

The query can be a natural language question or a search-style keyword string. Gemini may issue multiple searches to build a complete answer. Results include source URLs for verification.

Output is a synthesized summary (500-1500 words by default), not raw search results. Use maxResponseLength to adjust.

Agentic structured output: generate a JSON response conforming to a provided JSON Schema. Gemini runs inside workingDirectory with read_file and grep tools, so it can read files for context. Use for data extraction, classification, or any task needing machine-parseable output. The response is validated against the schema; isError is true if validation fails.

Retrieve a cached chunk from a previously chunked response. Large query and search responses may return only the first chunk plus a cacheKey. Use this tool with that cacheKey and a 1-based chunkIndex to fetch the remaining segments before the 10-minute in-memory cache expires.

Health check. Verifies gemini CLI is installed and authenticated, reports CLI version, auth status, configured models, and server version. Fast (~1s, no model call).