gemini-search-mcp
# gemini-search-mcp
<p align="center">
<img src="banner.png" width="700" alt="gemini-search-mcp">
</p>
<p align="center">
<img src="logo.png" width="120" alt="logo">
</p>
<p align="center">
MCP server for web search powered by Google AI Mode (Gemini). Free, unlimited, no API key.
</p>
## What is this
An MCP server that gives any AI agent (Claude, Cursor, Windsurf, etc.) the ability to search the web in real-time using Google's AI Mode — the same Gemini-powered search that lives in the "AI Mode" tab on Google Search.
Think of it as a free, unlimited alternative to Grok MCP / Tavily / SerpAPI, backed by Google's search index.
## Features
- **Free**: No API key, no subscription, no quota
- **Unlimited**: 60+ requests/min with zero rate limiting
- **Google quality**: Powered by Gemini + Google Search (grounded in real web results)
- **MCP native**: Works with Claude Desktop, Claude Code, Cursor, Windsurf, Cline
- **Also ships OpenAI API**: `/v1/chat/completions` for non-MCP clients
- **Fast**: ~1.5s average response time
## Quick Start
```bash
pip install -e .
# Optional: install the undetected-chromedriver backend for CAPTCHA probes.
pip install -e '.[undetected]'
```
### MCP Server (for AI agents)
```bash
gemini-search-mcp
```
### OpenAI-compatible API
```bash
gemini-search --port 8080
```
## MCP Integration
### Claude Code
```bash
claude mcp add gemini-search -- gemini-search-mcp
```
### Claude Desktop
Add to `claude_desktop_config.json`:
```json
{
"mcpServers": {
"gemini-search": {
"command": "gemini-search-mcp",
"args": [],
"env": {
"CDP_URL": "http://127.0.0.1:9222"
}
}
}
}
```
### Cursor / Windsurf
Same pattern — point to `gemini-search-mcp` as an stdio MCP server.
## MCP Tools
| Tool | Description |
|------|-------------|
| `web_search(query)` | Search the web and get a synthesized answer grounded in real-time results |
| `ask(prompt)` | General question — AI Mode auto-decides whether to search the web |
### Examples
```
web_search("latest AI regulation news 2026")
→ "The EU AI Act enforcement began on June 1, 2026, requiring..."
web_search("Bitcoin price today")
→ "As of June 30, 2026, Bitcoin is trading at $59,687 USD..."
ask("what is 1847 * 293")
→ "541171"
```
## OpenAI API Usage
```bash
curl http://localhost:8080/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model":"gemini-search","messages":[{"role":"user","content":"What happened in the news today?"}]}'
```
| Field | Value |
|-------|-------|
| Base URL | `http://localhost:8080/v1` |
| API Key | anything |
| Model | `gemini-search` |
## Environment Variables
| Variable | Default | Description |
|----------|---------|-------------|
| `CDP_URL` | (none) | Chrome DevTools URL. If set, connects to existing Chrome instead of launching one |
| `BROWSER_CHANNEL` | `chrome` | Browser to use: `chrome`, `msedge`, `chromium` |
| `HEADLESS` | `1` | Set to `0` to show browser window |
| `GEMINI_SEARCH_USER_DATA_DIR` | (none) | Persistent Chrome profile directory. Reuses cookies across runs and is not deleted on shutdown |
| `GEMINI_SEARCH_CDP_PORT` | `19250` | CDP port used for self-launched Chrome |
| `GEMINI_SEARCH_BROWSER_BACKEND` | `subprocess` | Browser launcher: `subprocess` or `undetected` |
| `GEMINI_SEARCH_PROXY_SERVER` | (none) | Chrome proxy server, e.g. `socks5://127.0.0.1:7897` |
| `GEMINI_SEARCH_CHROMEDRIVER` | (none) | Chromedriver executable used by the `undetected` backend |
## Persistent Chrome profile / CAPTCHA priming
If Google shows `/sorry/` CAPTCHA for a fresh temporary profile, prime a persistent profile once in a visible Chrome window, then reuse the same directory in headless mode:
```bash
# 1) Visible first run: solve CAPTCHA manually if Google asks.
gemini-search --no-headless --user-data-dir "$HOME/.local/share/gemini-search-mcp/chrome-profile"
# 2) Later runs: reuse the same cookies headlessly.
GEMINI_SEARCH_USER_DATA_DIR="$HOME/.local/share/gemini-search-mcp/chrome-profile" gemini-search
```
For Windows-side validation from WSL, run the probe with Windows Python through PowerShell so it launches Windows Chrome:
```powershell
$profile = Join-Path $env:TEMP 'gemini-search-mcp-persistent-profile'
python .\scripts\windows_chrome_profile_probe.py `
--profile-dir $profile `
--mode two-phase `
--out .\headless-reuse-result.json
```
Success evidence is `ok=true` and `stages.headless_reuse.captcha=false` in the JSON output.
## undetected-chromedriver CAPTCHA probe
When a normal Chrome subprocess gets a Google `/sorry/` CAPTCHA, install the optional backend and run the reusable probe against `google.com.hk`:
```bash
pip install -e '.[undetected]'
python scripts/uc_google_probe.py \
--proxy socks5://127.0.0.1:7897 \
--out-json uc-probe.json
```
Use the backend only when the probe reports `ok=true`, `captcha=false`, and `successful_for_engine_integration=true`.
```bash
gemini-search \
--browser-backend undetected \
--proxy-server socks5://127.0.0.1:7897 \
--chromedriver-path /path/to/chromedriver \
--no-headless
```
Observed on Windows Chrome for Testing 148 through Clash: headed UC passed (`captcha=false` and AI Mode tokens present), while headless UC hit Google `/sorry/`.
## How It Works
Google rate-limits by TLS fingerprint quality — not by IP. Automated HTTP clients (curl, requests, httpx) get throttled after a few requests. But a real Chrome browser's `fetch()` calls are trusted unconditionally.
This tool runs a single real Chrome tab and executes all queries as `fetch()` inside it over CDP, giving every request an authentic Chrome TLS/HTTP2 fingerprint. Google sees normal browser traffic and applies no rate limits. The optional `undetected` backend still uses the same CDP query path after launch.
```
Agent calls web_search("query")
→ Chrome Runtime.evaluate(fetch)
→ Google Search AI Mode (token extraction + folwr endpoint)
→ Parse answer from HTML response
→ Return to agent
```
## Comparison
| | gemini-search-mcp | Grok MCP | Tavily |
|---|---|---|---|
| Cost | **Free** | xAI API key ($) | API key ($) |
| Rate limit | **None** | API quota | API quota |
| Search backend | Google Search | Grok + web | Proprietary |
| Answer quality | Gemini synthesized | Grok synthesized | Extracted snippets |
| Setup | Chrome + CDP | API key | API key |
## Docker
```bash
docker compose up -d
```
## Requirements
- Python 3.10+
- Chrome, Edge, or Chromium
- Runtime dependencies from `pyproject.toml`
- Optional: `undetected-chromedriver` and `selenium` via `pip install -e '.[undetected]'`
## Limitations
- Requires Chrome/Edge/Chromium installed
- No conversation memory between requests
- Answer extraction relies on Google's DOM structure (may break on updates)
- Streaming is chunked, not per-token
## Acknowledgments
- [GenericAgent](https://github.com/lsdefine/GenericAgent) — 本项目核心开发依仗 GA 提供的 AI 能力
- [linux.do](https://linux.do) community
## License
MIT
TDQS
Scored across 2 tools
The two tools have distinct intentions: 'ask' for general questions that may use training data, and 'web_search' for explicit web-grounded answers. The description of 'ask' clarifies it is similar but flexible, reducing ambiguity. However, an agent might still be uncertain which to choose when a question could be answered from either source.
'ask' is a single verb, while 'web_search' is verb_noun with underscore. They do not follow a consistent pattern, which could be slightly confusing. However, both names are descriptive and readable.
With only 2 tools, the server feels thin for a search/QA domain. Typically a well-scoped server has 3-15 tools. However, for a specialized server focused on Google AI Mode, two tools (one for general questions and one for web search) might suffice, but it's on the lower end of acceptable.
The server covers two modes of AI-powered answering, but lacks a tool for pure training-data answers without web search, as 'ask' may still use web search. Users cannot force offline-only answers. Additionally, there is no tool for raw search results, only synthesized answers. This leaves notable gaps for a comprehensive search toolset.