nicobailon/pi-web-access v0.36.0

on GitHub

pi-web-access's search and fetch tools now work outside Pi. Run npx -y pi-web-access and Claude Code, Codex, Cursor, or any other MCP client can search the web, fetch pages, and check claims, using the same web-search.json and provider keys you already have. Docs sites that publish markdown now come back as clean markdown, and Firecrawl PDF costs can be capped. Several provider fixes also make everyday searching more dependable, especially with Brave prepaid keys and OpenAI usage limits.

Highlights

• Use pi-web-access's search and fetch tools from Claude Code, Codex, Cursor, and other MCP clients with npx -y pi-web-access.
• Docs sites that publish markdown versions of their pages now come back as clean markdown.
• Brave prepaid keys keep working after the first search, and OpenAI search shows the real error, such as a usage limit.
• Keep Firecrawl PDF costs down by reading only the first few pages of each PDF.
• New Keenable search provider that works without an API key.

Changelog

Highlights

• Use pi-web-access's search and fetch tools from Claude Code, Codex, Cursor, and other MCP clients with npx -y pi-web-access.
• Docs sites that publish markdown versions of their pages now come back as clean markdown.
• Brave prepaid keys keep working after the first search, and OpenAI search shows the real error, such as a usage limit.
• Keep Firecrawl PDF costs down by reading only the first few pages of each PDF.
• New Keenable search provider that works without an API key.

Added

• The search and fetch tools (web_search, fetch_content, get_search_content, and source_check) run as a local MCP server for Claude Code, Codex, Cursor, and other MCP clients. Start it with npx -y pi-web-access. It reads the same web-search.json and provider keys as the Pi extension and does not need Pi installed. Pi-only features such as the curator, summaries, and Kimi search are not available there. See "Use from other agents (MCP)" in the README. Thanks to @Avg8888 for issue #496.
• fetch_content asks servers for markdown first, so sites that publish markdown versions of pages (Cloudflare, Mintlify, and other docs hosts) return clean markdown directly. mode: "raw" still asks for the server's normal representation. Thanks to @erwinkramer for issue #495.
• web_search can restrict Exa results to a category such as news or research paper with the new category parameter. Other providers ignore it. Without an Exa API key, if Exa's filtered search is unavailable, the category is added to the query text instead. Thanks to @SuTang-vain for PR #493.
• New Keenable search provider, used only when you select it. It works without an API key through Keenable's public endpoint; set keenableApiKey or KEENABLE_API_KEY for higher limits. It supports domain and recency filters, routing, and the curator. Thanks to @ilya-bogin-keenable for PR #506.
• Set firecrawlPdfMaxPages to have Firecrawl read only the first N pages of each PDF, which caps per-page PDF billing on fetches and includeContent searches. Cut-off PDFs end with a note giving both page counts. Leaving it unset keeps the current behavior. Thanks to @Dangooy for issue #507.

Fixed

• Brave search with a prepaid (pay-as-you-go) key no longer fails every call after the first with "quota exhausted; estimated reset in ~702h". Prepaid plans report a monthly rate limit of 0, which was read as an exhausted quota. Thanks to @kk-code-lab for issue #501.
• OpenAI web search shows the server's real error when a response fails partway, such as a ChatGPT subscription usage limit, instead of "no parseable response output". Falling back to the next provider still works. Thanks to @ShinoharaHaruna for PR #505.
• web_search with includeContent keeps page content the search provider already returned and only fetches the pages it didn't cover. Before, it fetched every page again, and a failed fetch replaced usable content.
• web_search accepts provider, queries, and domainFilter lists that the model sent as a JSON string, such as provider: "[\"parallel-mcp\"]". Before, these calls were rejected before the search ran. Thanks to @advaitpaliwal for PR #491.
• fetch_content reports a tool error when its only URL failed or every URL in a batch failed. Before, Pi recorded these as successful calls even though the text said Error: .... Batches where some URLs succeed are unchanged. Thanks to @MDGChamomile for PR #504.
• get_search_content reads a fetched page without url or urlIndex when the fetch holds only one page. Before, calls like { responseId, findText } failed with "No URL specified" and the model had to retry with urlIndex: 0. Thanks to @j-koester for PR #494.
• When the model reads a long source_check result with get_search_content, each page now says how to get the next one, as search and fetch pages already did.
• Local video analysis through the Gemini Web cookie fallback no longer comes back as "garbled characters", because uploads now send the file's type and name. Thanks to @leonzyb for issue #508.