Web search for agents that won't bill your API budget

What it does

open-websearch gives agents and local tools the ability to search the live web and retrieve article content. It wraps multiple search engines—Bing, DuckDuckGo, Brave, Baidu, and several China-focused sources like CSDN and Juejin—into a single interface. Results come back structured with titles, URLs, and descriptions, and it can fetch full text from specific sites including GitHub READMEs and generic web pages.

The interesting bit

The project is deliberately API-key-free, relying on scraping and optional Playwright fallback for engines like Bing. It also ships a “skill” layer: a guided setup path that lets an agent self-diagnose whether the tool is installed, start the local daemon if needed, and pick the smallest working route. That is unusual agent UX—more autopilot than instruction manual.

Key highlights

• MCP server, CLI, and long-lived local HTTP daemon in one package
• Nine search engines, mix of global and China-specific sources
• Content extraction for CSDN, Juejin, GitHub READMEs, and generic pages
• Optional Playwright browser fallback for Bing; can reuse existing Chrome via CDP
• HTTP proxy support split cleanly between install-time and runtime

Caveats

• Playwright is no longer bundled; browser-enhanced search requires manual setup
• Some homepages and JS-heavy sites may not yield readable text through fetch-web
• Google support is listed in TODO but not yet implemented

Verdict

Worth a look if you are building agent workflows that need current web data but you would rather not manage another stack of API keys and rate limits. Less compelling if you already have paid search APIs wired in and do not need the China-specific sources.