mcp-gemini

# MCP Gemini Multimodal AI capabilities via Gemini MCP (lazy-loaded via ToolSearch). **CRITICAL**: Model is `gemini-3.1-flash-lite-preview`. The older `2.5-flash` and `2.0-flash` are DEPRECATED and return 404. ## Loading ``` ToolSearch: +gemini ``` ## MCP Tools | Task | Tool | When to Use | |------|------|-------------| | Chat | `mcp__gemini__gemini_chat` | Complex reasoning, long context | | Chat (fast) | `mcp__gemini__gemini_chat_flash` | Quick answers, simple tasks | | Analyze image | `mcp__gemini__gemini_analyze` | Screenshots, UI mockups, diagrams, CAPTCHAs | | Generate image | `mcp__gemini__gemini_image` | Create images from text prompts | | Fast image gen | `mcp__gemini__gemini_image_fast` | Quick image generation | | Edit image | `mcp__gemini__gemini_image_edit` | Modify existing images | | Code execution | `mcp__gemini__gemini_code_execute` | Run Python in sandboxed env | | Web search | `mcp__gemini__gemini_search` | Grounded search with citations | | Text-to-speech | `mcp__gemini__gemini_tts` | Generate audio from text | | Video analysis | `mcp__gemini__gemini_video` | Analyze video content | | Video status | `mcp__gemini__gemini_video_status` | Check video processing status | | Embed text | `mcp__gemini__gemini_embed` | Generate text embeddings | ## Common Patterns ### CAPTCHA Solving ``` 1. Screenshot the CAPTCHA: browser-agent screenshot -> /tmp/captcha.png 2. gemini_analyze(image_path="/tmp/captcha.png", prompt="Extract the text from this CAPTCHA") 3. Type the extracted text back into the form ``` ### Image Analysis (UI Review) ``` gemini_analyze( image_path="/tmp/screenshot.png", prompt="Analyze this UI screenshot. Identify layout issues, accessibility problems, and visual inconsistencies." ) ``` ### Grounded Web Search ``` gemini_search(query="latest React 19 features and breaking changes") ``` Returns results with citations -- use for current information beyond training data. ### Sandboxed Code Execution ``` gemini_code_execute(code="import pandas as pd\ndf = pd.DataFrame({'a': [1,2,3]})\nprint(df.describe())") ``` Runs Python with common data science libraries available. ### Video Analysis ``` gemini_video(video_url="https://youtube.com/watch?v=...", prompt="Summarize key points") gemini_video_status(operation_id="...") -- check if processing complete ``` Video analysis can be async -- use `video_status` to poll. ### Image Generation ``` gemini_image(prompt="A clean, modern SaaS dashboard with analytics cards") gemini_image_fast(prompt="Simple icon of a rocket ship") -- faster, lower quality gemini_image_edit(image_path="/tmp/original.png", prompt="Change the background to dark blue") ``` ## Decision Tree ``` Need current info? -> gemini_search (grounded, cited) Need to understand an image? -> gemini_analyze Need to generate an image? -> gemini_image (quality) or gemini_image_fast (speed) Need to run Python? -> gemini_code_execute Need video insights? -> gemini_video + gemini_video_status Need audio output? -> gemini_tts Need embeddings? -> gemini_embed General reasoning? -> gemini_chat (complex) or gemini_chat_flash (simple) ``` ## Gotchas - **Model version**: ALWAYS `gemini-3.1-flash-lite-preview` -- 2.5-flash and 2.0-flash return 404 - **Video is async**: `gemini_video` starts processing; poll with `gemini_video_status` - **Image paths**: Use absolute paths (`/tmp/screenshot.png`) not relative - **Search is grounded**: Results include citations -- use for factual claims - **Code sandbox**: Has pandas, numpy, etc. but no network access or file system persistence