pi-web-agent

@demigodmode/pi-web-agent is a Pi package for web research.

Most agent web tooling turns search, fetch, browser rendering, and synthesis into one blurry thing. This package keeps the model-facing surface simple instead: use web_explore for web research, and let it handle the lower-level work internally.

That lower-level work still matters. A search result is not the same as a page read. A plain HTTP read is not the same as rendering a page in a browser. But those are implementation details now, not separate tools the model has to juggle in normal use.

The contract

web_explore is the public research entrypoint.

Internally it can:

• read direct HTTP/HTTPS links from the prompt
• search for candidate sources
• fetch pages over HTTP
• escalate selected pages to headless rendering
• prefer forum/thread sources when the query asks for discussions
• rank official sources above weaker sources
• include community sources as practical context
• evaluate source quality, diversity, bot-checks, unreadable threads, and possible conflicts
• stop with a specific caveat when the evidence is thin or uneven

Preview shows up to three plain findings. Verbose also shows sources, evidence, and how each source was read, such as [HTTP], [Browser], or a special-content reader like [GitHub], [PDF], or [YouTube].

What you get right now

This project is still early, but it is usable.

Right now it gives you:

• one public web research tool: web_explore
• compact-by-default transcript output
• preview mode with up to three plain findings, plus verbose provenance and retrieval detail
• a /web-agent settings UI for presentation and backend config
• direct-link and forum/thread-aware research handling
• keyless public GitHub content plus PDF and YouTube readers that pull the real content behind those links
• questions about a pasted GitHub repo answered from a clone of its code, with commit-pinned citations
• Esc cancels a running research pass, and one page failing to load doesn't fail the run
• local-browser headless rendering with managed Chromium fallback
• optional hosted discovery through Brave, You.com, Exa, Tavily, Parallel, or a Google SERP endpoint without changing the public web_explore tool
• bounded research behavior that is willing to say when evidence was weak, narrow, blocked, or cautionary

Start here

• If you want to try it quickly, start with Getting started.
• If you want the install details, go to Install.
• If you want to understand the tool behavior first, go to Tools.
• If you want to change transcript output or backend config, go to Presentation and settings.
• If you want to work on the repo, go to Development.