@tduyng/pi-web

Web search and fetch for pi. The extension registers web_search and web_fetch, and registers the keyless Exa MCP server through pi's own MCP client: pi.registerMcpServer('exa', ...), exposure codemode, listed in /mcp. The tools call that server's two tools, web_search_exa and web_fetch_exa, through ctx.executeTool and carry no MCP SDK of their own. The server takes a 60 second request timeout, which covers the connect handshake, and pi retries only HTTP 408, 429 and 5xx: a handshake that overruns marks the server failed for the session and /mcp shows why.

Backends

web_search takes an array of up to six queries, one call each, and groups the results under the query that produced them. The primary backend is keyless Exa MCP, with DuckDuckGo HTML behind it. The Exa leg goes through the registered exa server, so an unconnected or --no-mcp session lands in the same DuckDuckGo fallback. x.com and twitter.com URLs go to Fixupx instead, which answers a page X refuses. It returns a post together with the whole long-form Article it links. An x.com/i/article/ URL carries an article id rather than a post id, so the reply names the post URL to fetch.

source: arxiv searches the arXiv Atom API instead, sorted by submission date, and returns abstracts with authors and dates. A miss falls back to the open web, because arXiv indexes preprints and nothing else.

source: hn searches Hacker News through Algolia's index, which needs no account, and returns each thread's points and comment count, the only measure of how much discussion a subject has. A comment hit is reported as the thread that holds it, with the comment's first 300 characters. A miss falls back to the open web.

Search returns excerpts, so the result ends with the ranked URLs to read. Without that list the model answers from a snippet.

Typed results

web_search declares an output schema, so a codemode script gets the answer as data in structuredContent.groups, one entry per query carrying title, url and snippet, plus published and authors where the backend has them, and points and comments on Hacker News. A script can compare values across results without parsing text. Exa's answer is prose: Title:, URL:, Published:, Author: and Highlights: blocks with a rule between them, which parseExa reads back into that shape, so the typed view holds for every backend. The model still receives the text, and the data is not sent to it a second time.

A backend that answers with nothing is a miss, and the next one runs. Before, an empty Exa answer ended the search and DuckDuckGo never got a turn.

DuckDuckGo queue

DuckDuckGo answers a burst of requests from one address with a bot check, and keeps that address locked out for 10 to 15 minutes, with each retry extending it. Measured on 2026-10-26: six requests at once got three pages of results and three challenge pages at 14 KB apiece, and the challenge arrives as HTTP 202 as often as a 403. Searches therefore run one at a time, two seconds apart. A bot check pauses the provider for five minutes, doubling to an hour while it keeps answering that way, and a paused search fails at once instead of waiting, so the batch reports that query as failed rather than holding the tool open.

Links in a fetched page

Exa's extraction returns plain text. Measured on 2026-09-27, arxiv.org/abs/1706.03762 came back with no links at all, so a citation cannot be read out of the page that cites it. Jina Reader returns Markdown with the links intact, 295 of them on a Hacker News thread, but it blocks a domain that hammered it: arxiv.org answered 403 until 2026-09-27. Exa answers first for that reason, and the links survive only when Jina does. Follow a citation by searching for the paper's title.

arXiv query form

arXiv ANDs nothing on its own. Measured on the live API: typed decision model returns 1,470,125 rows because the words are read as alternatives, and all:"typed decision model" returns 3 because the phrase is too strict. So a plain query has its words ANDed one at a time, which is the only form that ranks usefully. A query already written with arXiv field prefixes (ti:, cat:, au:, and the rest) is passed through untouched.

Reddit and X

Neither has a keyless search. Measured on 2026-09-27: reddit.com/search.json answers 403 on www, old and api, with a browser user agent and with a plain one, and X has no search without a key. DuckDuckGo's site: operator does return both, so a query that names the forum reaches those threads. The web backend finds them when a query names them: Hacker News and Reddit discussion of typed decision models returned a news.ycombinator.com host, while the bare query did not.

Quota

The keyless Exa MCP endpoint needs no account. Set EXA_API_KEY to send your own key. The extension reads it at load and puts it on the registered server URL as the exaApiKey query parameter. Exa's free plan includes 1,000 requests per month. A batch of six queries costs six of them, so they run three at a time.