AgentWebSearch

AgentWebSearch is an MCP server that exposes a set of web search and retrieval tools powered by real Chrome browser sessions. It runs as an MCP toolset that you can integrate into Claude Code, Cursor, LM Studio, or OpenClaw/Moltbot. Core capabilities include parallel web searches across Naver, Google, and Brave portals using real browser sessions, content fetching, and a specialized AgentCPM-Explore workflow for deeper, model-driven search planning. Tools available include web_search (parallel portal searches), fetch_urls (retrieve page content), smart_search (depth-controlled search with optional portal scoping), get_search_status and cancel_search (progress and partial results), and agentcpm (AgentCPM-Explore-based search with SGLang). You can also run the server in SSE mode for HTTP access. The workflow relies on non-headless Chrome instances to bypass bot detection and maintain login sessions, which enables persistent cookies and more natural search behavior.

Usage typically involves starting the required Chrome instances, then launching the MCP server and invoking the desired tools via your MCP client. For example, you can perform a fast, broad web search with smart_search, fetch the resulting pages with fetch_urls, or invoke agentcpm to execute a plan that uses the AgentCPM-Explore model to generate and execute specialized search queries. The server is designed to be integrated into Claude Code or OpenClaw workflows, allowing you to register agentwebsearch as an MCP tool and call its available endpoints from your agent pipelines.

Prerequisites:

• Python 3.8+ installed on your system
• Git available to clone the repository (or download the sources)
• Chrome/Chromium installed and accessible for remote debugging (CDP) support

Step-by-step:

1. Create a Python virtual environment and install dependencies

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

2. Install any optional browser tooling dependencies (if required by your environment)

• Ensure Chrome/Chromium is installed and accessible for chrome_launcher.py

3. Start the Chrome instances in debugging mode (3 instances by default)

python chrome_launcher.py start   # Start 3 Chrome instances

4. Run the MCP server

python mcp_server.py

5. (Optional) Run in SSE/HTTP mode

python mcp_server.py --sse --port 8902

6. Verify status and stop when needed

python chrome_launcher.py status
python chrome_launcher.py stop

Notes:

• The repository expects Chrome to be launched with a debugging port so that the MCP server can control it via CDP.
• If you want persistent login sessions, configure the chrome_launcher.py to use non-temporary profiles as described in the README.

Tips and caveats:

• The AgentWebSearch MCP tool relies on real Chrome sessions; ensure Chrome is properly installed and accessible from the environment running the MCP server.
• If you encounter CAPTCHA or login issues, verify that the Chrome instances have stable sessions and that profiles are not being cleared between runs.
• For best results with agentcpm, ensure SGLang (AgentCPM-Explore) is available if you plan to use the agentcpm tool; it assumes a running SGLang server on port 30001.
• When integrating with MCP clients, register the agentwebsearch tool under the mcpServers entry (as shown in the Quick Start) and reference the tool by the name agentwebsearch.
• If you need parallel searches across portals, the tool supports starting multiple Chrome instances in parallel and coordinating results via the MCP interface.