For an agent running continuously, a metered search API is a failure mode; this one has no key and no quota because it queries the engines directly and re-ranks locally. It is built for the Chinese network in particular — the default engines are reachable there without a proxy — and the tool descriptions themselves carry a lot of hard-won advice about queries that tokenizers mangle.
A local search server with two tools: web_search queries Baidu, Sogou, Bing CN and Bing International together and merges the results, and web_fetch reads a page in full.
- web_search returns ranked titles, URLs and snippets, with a topic switch between general web results and news
- A freshness filter for the past day, week, month or year, meant to be used instead of putting dates in the query text, which skews the ranking
- Per-call domain allow-lists and block-lists, so a noisy source can be dropped without a permanent configuration change
- Results from all four engines are merged and re-ranked by embedding similarity to the query rather than simply concatenated
- web_fetch returns clean extracted text with navigation, ads and boilerplate removed
- Extraction handles server-rendered modern frameworks and pulls structured data out of pages that plain scraping returns nothing for
- Results unrelated to the query are detected and flagged, and four independent engines give the model something to cross-check against
- Both tools carry detailed guidance in their own descriptions on reformulating a query when the results come back as noise
Python, installed from the package with a one-time setup step, or the published container image. The embedding model downloads once through a mirror on first run and is then served from local cache. No account, no API key and no usage limit — the pipeline runs on your machine.
One command — pip install searchpin && searchpin-setup