Content
# Dokobot MCP Server
Dokobot MCP Server - Real browser parallel web search capabilities for AI agents
> **Note**: This project is an MCP server wrapper for [@dokobot/cli](https://github.com/dokobot/dokobot). Dokobot CLI tool is developed and maintained by the [Dokobot](https://dokobot.ai) team.
## Features
- Intelligent search engine routing: Chinese queries use Google/Bing/Baidu, English queries use Google/Bing/DuckDuckGo
- True parallel search: Multiple search strings executed simultaneously, automatically assigned engines
- URL scoring and sorting: Cross-engine consensus + SERP position + authoritative domain name bonus
- Automatic fallback: Page read failure automatically tries the next candidate
- 5-minute cache: Avoids duplicate requests to the same page
- Concurrency limit: Up to 4 browser tabs concurrent
## Prerequisites
1. Node.js >= 18
2. Global installation of [@dokobot/cli](https://www.npmjs.com/package/@dokobot/cli)
```bash
npm install -g @dokobot/cli
```
3. Chrome browser + Dokobot extension
4. Local bridge installed
```bash
dokobot install-bridge
```
## Installation
```bash
npm install -g dokobot-mcp-server
```
## Usage in Astrbot
Edit `~/.astrbot/data/mcp_server.json`:
```json
{
"mcpServers": {
"dokobot": {
"command": "node",
"args": ["/path/to/global/node_modules/dokobot-mcp-server/dokobot-mcp.js"],
"env": {}
}
}
}
```
Or use npx (no global installation required):
```json
{
"mcpServers": {
"dokobot": {
"command": "npx",
"args": ["dokobot-mcp-server"],
"env": {}
}
}
}
```
Restart Astrbot, and the MCP tool is ready to use.
## Tool List
### 1. `quick_search`
Lightweight search, defaults to reading 2 most valuable pages
```javascript
{
"query": "Trump 2026 latest news",
"engine": "google", // Optional: google/bing/baidu/duckduckgo/sogou
"site": "bbc.com", // Optional: limit to a specific website
"read_pages": 2, // 0-4, 0 = only read search results page
"max_chars": 8000
}
```
### 2. `deep_research`
In-depth search, multiple search strings in parallel + automatic selection of most valuable pages for in-depth reading
```javascript
{
"query": "Trump visit China",
"keywords": [
"Trump visit China 2026 US-China relations latest tariffs trade talks results foreign ministry statement",
"Trump China Xi meeting trade visit tariff policy 2026",
"US-China high-level meeting semiconductor chip export control entity list countermeasures"
],
"read_pages": 4 // 0-8
}
```
### 3. `read_url`
Read a specific webpage (supports Bilibili/YouTube video pages)
```javascript
{
"url": "https://example.com/article",
"screens": 4 // Optional: for long pages, set 4-8
}
```
## Working Principle
1. **Engine Routing**: Detect query language, automatically assign the most suitable engine
2. **Parallel Search**: Multiple search strings executed simultaneously, up to 4 concurrent
3. **URL Scoring**:
- SERP position weight (higher for top positions)
- Cross-engine consensus (bonus for multiple engines returning the same result)
- Authoritative domain name bonus (Wikipedia/GitHub/Zhihu/etc.)
4. **Fallback**: Page read failure automatically tries the next candidate
5. **Cache Optimization**: Same URL reused within 5 minutes
## Configuration
The server automatically looks for `@dokobot/cli`:
- Prioritizes: `C:\Users\{user}\AppData\Roaming\npm\node_modules\@dokobot\cli\dist\cli\bin\dokobot.js`
- Falls back to: `dokobot` command in PATH
## License
MIT
