Pi Web Access

Pi Web Access is an extension for the Pi coding agent that adds web search, URL content extraction, and video understanding. It registers four tools: web_search, fetch_content, get_search_content, and source_check. With no API keys configured, search runs through Exa MCP, and a Pi session signed in with a Codex subscription can reuse that auth for OpenAI web search. The same search and fetch tools also run as a local stdio MCP server for Claude Code, Codex, Cursor, and other MCP clients, without the Pi-only features (the curator, summaries, Kimi search, answer mode, and video prompts).

Every capability has a fallback chain. A search tries a configured self-hosted SearXNG endpoint first, then subscription-backed OpenAI search when the active Pi model uses a ChatGPT subscription, then Exa, OpenAI, Brave, Parallel, TinyFish, Search1API, Searchinfinity, Querit, Tavily, Firecrawl, Jina, SERPdive, Kagi, Bocha, Ollama, Perplexity, and Gemini. Some providers (xAI, Mistral, Kimi, Bright Data, SerpApi, Serper, and the other paid SERP APIs) are explicit-only: auto and all never pick them, so one query cannot spend a subscription quota unless you name the provider. provider: "all" runs a query against every eligible provider at once and returns one result card per provider.

fetch_content routes by URL type. GitHub repository URLs are cloned locally so the agent reads real file contents at a local path (repositories over 350 MB get an API-based view instead), and pull request and issue URLs render as one markdown document with status, checks, reviews, files, and comments. YouTube URLs and local video files go through Gemini for transcripts, visual descriptions, and chapter markers, with frame extraction at exact timestamps. PDFs convert to Markdown through Datalab (Marker), Gemini, or local pdf.js text extraction. HTML pages run through Readability, then configured Firecrawl or Crawl4AI, and hosted fallbacks only when you opt in. Fetched content lands in a bounded local cache and pages back through get_search_content.

Inside Pi, an optional curator opens a browser page where you review, add, and select search results and approve a summary before it returns to the agent. The /websearch, /curator, /search, and /google-account commands and a live activity monitor cover the interactive side. Tool activation can be dynamic: a session starts with one compact web_enable tool, and the configured web tools appear once the model calls it.