Query Google AI (Gemini models) with a prompt. This tool operates as an intelligent agent with multi-turn execution capabilities. The agent can automatically use available tools (web fetching, external MCP servers) to gather information and provide comprehensive answers. Supports multi-turn conversations when sessionId is provided. Supports multimodal inputs (images, audio, video, documents) via the optional 'parts' parameter.

Search for information using Gemini. Returns a list of relevant search results. Follows OpenAI MCP specification for search tools.

Fetch the full contents of a search result document by its ID. Follows OpenAI MCP specification for fetch tools.

Generate images using Gemini's native image generation (Nano Banana). Supports gemini-3-pro-image, gemini-3.1-flash-image, gemini-3.1-flash-lite-image, and gemini-2.5-flash-image models. gemini-3.1-flash-lite-image (Nano Banana 2 Lite) is the fast, low-cost GA tier: 1K output only, standard aspect ratios (no 1:4/1:8/4:1/8:1), no thinkingLevel, up to 14 reference images. gemini-3.1-flash-image is required for 0.5K and 1:4/1:8/4:1/8:1 ratios. gemini-2.5-flash-image (legacy, retires 2026-10-02) supports at most 3 reference images and does not support imageSize. Reference images use imagePaths and must be PNG (.png), JPEG (.jpg/.jpeg), WEBP (.webp), HEIC (.heic), HEIF (.heif). Audio and video reference files are not accepted by generate_image. Images are saved to /root/Pictures/gemini-generated and returned as base64.

Generate speech audio using Gemini TTS models. Supports single-speaker and two-speaker TTS. Input is text-only; audio, image, and video reference files are not accepted. TTS has a 32k-token context limit and does not support streaming. Audio is saved to /root/Music/gemini-generated/speech and returned as MCP audio content.

Generate music using Lyria models. Supports lyria-3-clip-preview fixed 30-second clips and lyria-3-pro-preview full songs up to 184 seconds. Lyria 3 supports one clip per prompt; language directions follow the model-card set: English, German, Spanish, French, Hindi, Japanese, Korean, Portuguese. Gemini API/AI Studio mode supports 44.1 kHz stereo audio/mp3 output, and audio/wav only for lyria-3-pro-preview. Negative prompting is not supported. Lyria 3 accepts text prompts and optional imagePaths (PNG (.png), JPEG (.jpg/.jpeg), WEBP (.webp), HEIC (.heic), HEIF (.heif)); audio/video reference files are not accepted. Audio is saved to /root/Music/gemini-generated/music and returned as MCP audio content.

Start video generation using Google's Veo models. Returns an operationId immediately. Video generation typically takes 1-3 minutes. Use check_video with the operationId to poll for completion and download results. Recommended polling interval: 30 seconds. Supports text-to-video, image-to-video (with imagePath), interpolation (imagePath + lastFramePath), reference images (referenceImagePaths, max 3, Veo 3.1 only), and Veo video extension (videoPath). Image source file types: PNG (.png), JPEG (.jpg/.jpeg), WEBP (.webp). videoPath must be MP4 (.mp4) from a previous Veo 720p generation. Audio file references are not supported; describe dialogue, SFX, and ambience in the prompt instead.

Check the status of a video generation operation. Returns status: 'running' (still generating), 'completed' (with saved file paths), or 'failed' (with error). If status is 'running', wait ~30 seconds before checking again. Completed videos are saved to /root/Videos/gemini-generated.

Generate or conversationally edit short videos with Gemini Omni Flash (gemini-omni-flash-preview). This is a NON-Veo model on the Google AI Studio (Gemini API) backend and does NOT use generate_video/check_video: it returns the finished video synchronously in one call (no operationId polling). Two paths: (1) ONESHOT generation — text-to-video, or image/reference-to-video via imagePaths (max 7); (2) INTERACTIVE editing — set previousInteractionId to an id returned by a prior call to edit that video with a natural-language instruction (no image re-upload; chain up to 3 sequential edits). Constraints: 720p output only; aspect ratio 16:9 or 9:16; clips run a few seconds (steer pacing/timing within the prompt — duration is not a parameter); a synced audio track is generated automatically (audio reference inputs are not accepted — describe dialogue/SFX/ambience in the prompt). Image source file types: PNG (.png), JPEG (.jpg/.jpeg), WEBP (.webp). The response includes interactionId (pass it back as previousInteractionId to edit) and the saved file path. Videos are saved to /root/Videos/gemini-generated.

AI-assisted reference search: answer a question from live web sources using Gemini's Google Search grounding, and return organized citations. Unlike the OpenAI-spec 'search'/'fetch' connector tools, this composes a synthesized answer AND returns the source links plus claim->source supports (citations) in one call. Returns: answer (synthesized text), citations (deduped {index,title,uri,domain} sources), supports (answer segments mapped to citation indices with confidence scores), searchQueries (the queries the model actually ran), and searchSuggestionsHtml (Google's required Search Suggestions markup to display alongside the answer). Search-scope tuning is backend-specific: Vertex AI supports excludeDomains (skip up to 2000 domains) and blockingConfidence (block risky/low-quality sites); Google AI Studio supports timeRange (restrict to a publish-time window) and grounding on explicit urls via URL context (Gemini API only, not Vertex AI). Both backends support includeImages.