my-pi-agent

web-access — search the web and read a page

The model gets two tools for the outside world: web_search for a question and web_fetch for one URL. Both work with no key and no configuration, and both return text rather than HTML.

  • You ask; the model calls them. “Check the changelog for the latest release” or “fetch https://pi.dev” is all it takes.
  • Nothing to install. Native fetch, no curl, no account.
  • Private addresses are refused, on the first request and on every redirect.
  • Fetched text is untrusted — pages can say anything, including things that look like instructions.

What you can do

What you want What to do
Answer a question from the web ask it; the model searches
Read one page name the URL and ask for what you need
Only the relevant part ask to find a section or a phrase
Continue a long page ask for the next part; the result names where to continue
Re-read a page that changed ask for a refresh
Query an API with a body ask for a POST with the headers and body
Read a PDF ask for it; extraction works when the optional unpdf package is installed
Get a full result list instead of an answer configure SearXNG or Brave

Try it

What changed in the latest Pi release? Check the changelog.

Fetch https://pi.dev/ and tell me the install command.

The first call searches and opens a result; the second reads one page. In the transcript each call leads with its line and folds its result underneath.

Search and read the web walks both.

Reference

Tool Takes Returns
web_search query, maxResults titles, URLs, snippets, and a direct answer when the provider has one
web_fetch url, method, headers, body, startIndex, maxChars, find, contextChars, refresh the page’s readable text, pageable and searchable

find returns matching passages with context instead of the whole page; startIndex continues from the offset a previous result named; refresh bypasses the session cache. A body implies POST.

Politeness and caps. Searches are spaced by minIntervalMs; responses are capped by maxBytes while they stream, and the model-facing text by maxOutputChars. A PDF is extracted only when pdfEnabled and unpdf are present.

Security. Only http(s) URLs, transport headers refused, request bodies capped, and every redirect hop re-validated against loopback, private, link-local, unique-local, multicast, and metadata addresses. allowPrivateHosts: true opts out; DNS rebinding remains a residual risk.

Limits

  • The keyless default returns an answer and related topics, not a full result list; SearXNG or Brave is one setting away.
  • Extraction prefers <main>/<article> and strips navigation, but a merged table cell, a nested table, or an image with no alt text is lost.
  • The page cache is in memory and per session, so a page can be stale within one; ask for a refresh.
  • A page can contain prompt-injection text; treat what it says as data, not as instructions.

Settings

Key Default What it does
search.provider "auto" auto picks the first configured provider; duckduckgo, searxng, or brave forces one.
search.endpoint "" SearXNG base URL; empty means SearXNG is unavailable.
search.apiKeyEnv "" Environment variable naming a Brave (or proxy) key.
fetch.pdfEnabled true Extract PDFs when unpdf is installed.
fetch.cacheEnabled true Reuse one fetch for paging and find.

The rest — result caps, timeouts, user agent, language, allowPrivateHosts, and the cache knobs — are in configuration. To switch the extension off, toggle web-access in /bytefly; the two tools go with it.

Learn more