noapi-google-search-mcp

vincentkaufmann@protonmail.com

If you want to support noapi-google-mcp or gpt-oss-20B/120B-Vision and other open source models in the pipeline please add to my GPU compute fund: around $3000 are needed.

If you find this useful, consider supporting continued development and new features.
BTC: 16DT4AHemLyn7C6P116YepjY518gu9wUUH
ETH: 0x7287D1F9c77832cFF246937af0443622bFdACD04

38 tools. Zero API keys. Give any local LLM real Google search, live feeds, vision, OCR, and full video understanding.

An MCP server that turns your local LLM into a fully connected assistant. Real Google results, live news and social feeds, reverse image search, offline OCR, YouTube transcription and clip extraction — all running locally through headless Chromium and open-source ML models. No API keys, no usage limits, no cloud dependency.

Works with LM Studio, Claude Desktop, OpenClaw, Ollama, and any MCP-compatible client.

What's New in v0.3.1 — Anti-Bot Detection & Neural Net CAPTCHA Solver

Google started catching headless Chromium browsers and serving CAPTCHAs instead of search results (#2). Fixed it properly and then added a neural net CAPTCHA solver just for the hell of it.

Stealth patches:

• JavaScript injection before every page load patches navigator.webdriver, fake browser plugins, chrome.runtime, and removes Playwright fingerprints
• Fixed User-Agent mismatch (was aarch64, now matches the actual Chromium binary)
• Cookie persistence between sessions so Google sees a returning browser, not a fresh bot
• Human-like random delays between page interactions
• Clear error messages when rate-limited instead of cryptic timeouts

Neural net CAPTCHA solver (MobileNetV2 + OpenCV):

• When Google serves a reCAPTCHA image challenge ("Select all images with traffic lights"), the server now attempts to solve it automatically
• Uses MobileNetV2 (ImageNet classifier, ~13MB ONNX model, auto-downloads on first encounter)
• OpenCV splits the CAPTCHA grid into cells, each cell is classified by the neural net and matched against the prompt
• Covers 22 common CAPTCHA categories: traffic lights, buses, bicycles, motorcycles, cars, bridges, boats, trains, trucks, fire hydrants, parking meters, tractors, and more
• Human-like mouse movement for clicking cells and the verify button
• Falls back gracefully if the challenge type isn't supported

Will it solve every CAPTCHA? No. But it'll get through some of them, and honestly we mostly built it because we could.

What's New in v0.3.0 — 16 New Tools

YouTube RAG — Subscribe, Transcribe, Search, Clip

A full local retrieval-augmented pipeline for YouTube. Subscribe to any channel, and when you check feeds the server automatically downloads the audio, transcribes it with Whisper, and stores the full transcript in SQLite with FTS5 full-text search. Every word from every video becomes searchable — ask your LLM a question and it pulls the answer from the right video.

Found something interesting? Ask the LLM to extract that segment as a clip. It reads the transcript, finds the exact timestamps, and cuts the video for you. This is the beginning of AI/LLM-powered video editing — your LLM understands the content and cuts clips based on what's being said, not just timestamps.

How it works:

1. subscribe to a YouTube channel — by handle (@3Blue1Brown), URL, or channel ID
2. check_feeds — new videos are fetched and auto-transcribed locally with faster-whisper. Transcripts are written into the database for instant search. Up to 5 videos per check, cached so nothing gets re-downloaded.
3. search_feeds — full-text search ac

... View full README on GitHub