Hi HN. CandleFeed is a small API for crypto derivatives data (funding, open interest, liquidations). We wanted to test execution ideas against what the Binance book looked like, and commercial tick data cost more than we could justify for one venue, so in June we started recording it.

Three nodes record USD-M depth@100ms diff stream, the trade stream, and a 1,000-level REST snapshot every 5 minutes. Two are in Tokyo on different providers and one is in Singapore. A nightly job merges them: an update from a second node only goes in where Binance's update-id chain (pu equals the previous u) stays continuous, and where nodes overlap, their copies have to match, or the day is held back. Anything all three missed goes into a public gap log. The largest trade gap so far: on 2026-08-19 at about 15:28 UTC, all three nodes lost the same 32,153 trades across ten symbols, about 4 to 5.5 seconds' worth depending on the symbol.

A surprise: on 2026-09-25, 848 of the 861 BTCUSDT snapshots our three nodes took landed partway through a diff event. (That count comes from the nodes' raw copies; the published snapshot file carries one node's snapshots.) Replay has to apply that "straddling" event in full, so we tag every snapshot clean, straddle or in_gap.

Try it without an account: `curl https://candlefeed.ai/api/v1/l2/sample` lists the two open days, BTCUSDT and DOTUSDT book data for 2026-10-01. In Python, after `pip install "candlefeed[l2]>=0.3.2"`, run `from candlefeed import CandleFeed` and then `CandleFeed(public=True).download_l2_sample("DOTUSDT", "2026-10-01", "data/")`. That downloads the DOT day (about 100 MB) and checks every file's SHA-256. Sample files, with or without an account, are for internal use under our Terms (https://candlefeed.ai/terms, §5.3), so please don't re-host them. With a free account, the 1st of each month is free for every symbol (10 GiB a month). Pro is $149/mo for the full archive.

Our MCP server, candlefeed-mcp (`pip install candlefeed-mcp`, https://pypi.org/project/candlefeed-mcp/), lets Claude Code or Cursor pull a day and rebuild the book from a prompt; it needs a free API key. There's also a notebook that does it in plain Python on the 2026-10-01 BTCUSDT sample. It uses a free API key too; for a no-account run, swap its download call for `download_l2_sample`: https://candlefeed.ai/examples/l2-free-sample-day/

What's weak: one venue, 25 symbols (10 with history from June 2026, 15 from late September), and we can't backfill what we didn't record. It's also the 100ms stream, not Binance's fastest.

Show HN: Binance futures order book history from 3 nodes, with a public gap log