All stories
Pi Web Access Ecosystem Community reports

Pi Web Access adds three search routes and trims MCP overhead

Version 0.46.0 adds explicit-only Cloudflare, SearchApi and Tabstack search, opt-in hosted page readers, and fixes for PDF text and proxy persistence.

Sources (2)

Why now The maintainer published v0.46.0 on October 29 at 04:27 UTC, establishing a specific new release rather than just repository activity.

Pi Web Access 0.46.0 gives coding agents more choices for searching and reading the web while reducing the space its MCP tool definitions take in model requests. The October 29 release adds Cloudflare Web Search, Google through SearchApi, and Tabstack Research. All three are explicit choices, according to the maintainer: ordinary auto and all searches do not silently select these new paid routes.

More routes, with different costs and permissions

Cloudflare search requires a Cloudflare API key and account ID, using AI Gateway credits or a provider key stored on the gateway. A Cloudflare credential saved through Pi’s model login is not reused for this route. SearchApi requires its own key and returns Google results; Tabstack requires its own key and returns a cited research answer rather than a conventional list of links.

The release also adds optional Exa and Tabstack page-reading routes for sites that refuse a direct fetch. These require adding the chosen reader to fetchRouting.providers and enabling fetchRouting.allowRemoteHostedProviders. For Exa, the notes explicitly say the page URL is sent to the service and each page read is billed. This is a separate permission from configuring a search key.

An optional curl fetch path can use curl-impersonate to connect like Chrome. The maintainer reports that readable pages rose from 18 to about 31 in a test of 68 mostly bot-protected pages. That is an attributed test of a particular sample, not a promise to read every blocked site or bypass authentication.

Smaller schemas and useful fixes

The release notes put MCP tool-definition size at roughly 7,000 characters before and 3,500 after, estimating about 1,750 versus 900 tokens. This concerns tool-schema overhead in each model request, not a halving of the entire conversation’s token use.

Other fixes return a PDF’s extracted text instead of merely its saved-file notice, preserve a configured proxy after Pi extension reloads, and distinguish rate-limited free search from a missing provider. The release also labels OpenAI and xAI search-answer snippets as provider summaries rather than source-page quotations.

Pi users install the MIT-licensed extension with pi install npm:pi-web-access; the documentation requires Pi 0.37.3 or newer. Claude Code, Codex, Cursor and other MCP clients can instead launch its local stdio server with npx -y pi-web-access, requiring Node.js 22.19 or newer and no Pi installation. That MCP path omits Pi-only features, including ChatGPT-sign-in search and video prompts or frames. Existing zero-config Exa MCP search remains an alternative to paid provider keys. AITrending has reviewed the English release and documentation but has not reproduced the fetching or token measurements.