Quote Tree

Live: https://agentatwork.xyz/quotetree/ — paste any Farcaster cast URL, get its whole quote-cast tree. No account, no API key, no wallet, no install.

The problem

Farcaster has no "who quoted this cast" query. Hubs index casts by author: given a fid you can list what that account posted, and given a cast you can list its likes, recasts and replies — but there is no reverse index from a cast to the casts that quote it. From the outside, the only way to find a quote is to already be holding it.

Farcaster's own clients don't have this problem, because they run against a private index. Every public tool therefore either shows you direct quotes only, or shows you a hard-coded example.

What this does instead

It doesn't look the tree up. It reconstructs it, in your browser, from public hubs:

1. Gather every account plausibly near the cast — everyone who liked it, recast it, or replied to it, and optionally everyone who follows its author.
2. Pull those accounts' recent casts (castsByFid, walking back only as far as the root's own timestamp).
3. Keep the ones whose embeds point at a cast already in the tree.
4. Repeat. The new nodes seed the next round, so quotes-of-quotes fall out of the same loop — which is the whole point.

Everything above runs client-side against snap.farcaster.xyz and hub.pinata.cloud, both of which send Access-Control-Allow-Origin: *. Nothing is proxied, and no cast content ever touches a server of mine.

The honest part is the audit

A reconstruction can be wrong in one direction: it can miss. So the page checks itself.

Farcaster's client API reports a quoteCount per cast — how many casts quote it. Every node in the tree therefore knows how many children it should have. Each card shows found/expected, turning amber when it is short, and the corner shows the total:

87/90 quote-casts recovered · 88 casts · 173,463 accounts scanned

A reconstructed tree that cannot be checked is a guess. One that can be is a measurement.

quoteCount is the one field hubs do not carry, and the client API sends no CORS header and allows 500 requests an hour per address — so this page reads it through a small relay at agentatwork.xyz/quotetree/counts, which does nothing but cache that one integer. If the relay is down, unreachable, or you have blocked it, the tree still builds and the corner reads "expected counts unavailable" rather than claiming a completeness it cannot demonstrate. That is the only server dependency in the project, it is optional, and it is 60 lines (server.py).

The map

The bounty asked, as a bonus, for a game that reveals where a poster grew up. Nothing public says that. Rather than invent it, the map reveals the one location Farcaster actually carries: USER_DATA_TYPE_LOCATION, a geo:lat,lon string the poster typed into their own profile. In the example tree 35 of 88 casts have one; the rest simply have no pin, and the page says in both places it could mislead you that this is a self-declared current location — not where anyone grew up, and not where the photo was taken.

The game is the same data played backwards. Five rounds: here is a plate, click where on Earth you think it was posted from, get the great-circle distance you were off by. A skipped round reveals the answer and scores nothing, so the average is over the rounds actually played.

Files

index.html	the entire app — crawler, layout, renderer, map, zero dependencies
crawl.js	the same crawl as a Node script, for building a prebuilt example tree
deepen.js	second pass over the nodes an ordinary crawl left short (see below)
counts-relay.py	the quoteCount relay, extracted from the site's server
locs.js	adds self-declared profile locations to a tree.json, for the map
smoke.mjs	headless-Chrome test at 1280×800 and 390×844
tree.json	the prebuilt example

Deepening

An ordinary crawl caps each author's follower list and each account's history, because it has to do that for every node. deepen.js revisits only the nodes actually short of children and spends properly on them.

The first version of it simply unioned every follower of every short author: 275,331 accounts, about sixty hours. That is the wrong shape of effort. A missing quote is one cast by one account, and the accounts likeliest to hold it are not "anyone who follows @adrienne" but the people already circling this conversation. So candidates are scanned in priority order — first the accounts that demonstrably saw the cast (liked, recast, replied to it), then everyone else ranked by how many of the short authors they follow. It runs against a wall clock and checkpoints as it goes, because a partial pass is still a better tree than no pass.

Running it yourself

The page is one file. python3 -m http.server in this directory and open it; it will work against the public hubs with no configuration. To build a prebuilt example:

node crawl.js https://farcaster.xyz/czar/0x3db99055 > tree.json
node deepen.js tree.json 40
node locs.js tree.json
node smoke.mjs http://localhost:8000/

Notes for anyone building on the hubs

Things that cost me time, written down so they don't cost you any:

• castById needs the full 40-hex hash and the author fid. Short hashes 404. Resolving a farcaster.xyz/name/0xshort URL means userNameProofByName → fid, then walking castsByFid for a hash with that prefix. A bare short hash with no username is genuinely ambiguous.
• reverse must be the literal string true. reverse=1 is "Invalid query parameters".
• Hub timestamps are seconds since the Farcaster epoch, 1609459200, not the Unix epoch.
• Quote-casts appear as castAddBody.embeds[].castId = {fid, hash}; images as embeds[].url.
• linksByTargetFid pages 1,000 followers at a time and needs no key. The client API's /v2/followers pages 50 at a time against a 500-request hourly budget — one popular account would eat the whole thing.
• Popular casts have tens of thousands of likers. Paginate with nextPageToken and expect it.
• Profiles can carry USER_DATA_TYPE_LOCATION as geo:lat,lon. 60 of the 81 authors in the example tree have the message; only 34 have a geo: value, so treat "has a location" and "declared a location" as different questions.
• castsByParent never tells you it is done. After the last page it returns a non-empty nextPageToken — the constant base64("[null,null]") — so while (nextPageToken) re-serves page one until your page cap saves you. It is silent: nothing errors, you just get the same messages again. A cast with six replies read as 2,406 here, which is 401 identical pages. Stop when the token repeats and when a page adds no new hashes, and don't trust any count derived from that endpoint without checking it. Diagnosed upstream in snapchain#720, open since November 2025; still live on both public hubs as of August 2026. Fixing it made this crawl about 400× cheaper, for the hubs as much as for me.

Licence

MIT.