DepthFeed

What it is

An archive of the thing exchanges do not keep: the complete bid and ask ladder, both sides and every level, recorded at every change rather than sampled on a clock. Polymarket is captured off its CLOB WebSocket, Kalshi by paced full-depth REST polling with up to 100 levels a side, and Limitless through its push socket. The scope is narrow on purpose — crypto up-or-down markets on those three venues plus Binance Wallet's Predict.fun-backed markets, across seven assets in 5-minute to 24-hour windows — and a separate sports surface sits beside it.

The read surface is small and predictable. GET /v3/{coin}/markets discovers, then /v3/{coin}/markets/{id}/snapshots returns recorded books; /v3/bars serves OHLCV for US equities, ETFs and Binance spot and perpetual pairs inside the same plan history window; /v3/sports/markets is the sports catalogue; /v3/whoami echoes the key's plan and limits. Live books come from wss://api.depthfeed.com/v3/stream carrying the same JSON objects as the REST snapshots. There is also a remote MCP server at api.depthfeed.com/mcp.

Two response mechanics are worth understanding before the first query, because they decide whether a backtest is honest. ?interval= selects rather than interpolates: a bucket appears only if a book was actually captured in it. ?fill=ffill gives a dense grid instead, carrying the last captured book forward into empty buckets, and marks each row with filled and as_of_ts so staleness is measurable rather than assumed. Any response that used interval carries a resolution block reporting buckets in range, buckets filled, rows returned and whether the window is complete.

Every snapshot is stamped with exchange and receive times in epoch milliseconds and carries a nearest-preceding Binance reference price where one is available, which is what makes the market's own 0-1 price joinable to the underlying it settles against.

Availability

No account needed to look, no identity check to buy, and no geographic gate observed: the keyless demo endpoint answered this machine on 19 September 2026. Paid plans are billed monthly in US dollars through Stripe, and the vendor states that a paid plan with no successful billable API call since the period began is refunded automatically.

The operator describes itself as an independent research project with no affiliation to or endorsement from any of the venues it records. That is a licensing fact as much as a modesty one — what a venue's own terms permit you to do with its prices is not settled by a subscription here.

Pricing

Four plans: Explorer free, Quant at 29 USD a month, Research at 99 and Desk at 249. They are priced on one axis above all others — how far back you can look. Explorer gets seven days and a capped sample, Quant thirty days, Research ninety, and only Desk gets the full archive with no rolling cap.

Three other gates move with the price. The raw tick tape — every book change in venue sequence — is Desk only. Snapshot interval is 60 seconds minimum on Explorer and arbitrary on every paid plan. And Explorer is limited to a Polymarket and Kalshi sample rather than every venue, one request a second, and a single BTC live subscription.

REST admission runs 1, 25, 50 and 100 requests a second across the four plans. Those are general admission numbers, not archive delivery rates: historical queries have their own account-level concurrency, with one request in flight on the first three plans and two on Desk, and two distinct 429s — one saying your account is over its own allowance, one saying the shared archive pool is saturated and that reducing your own parallelism will not help.

The only thing that can cost extra is an optional Research Compute add-on, off by default. The "AI execution credits" on each plan pay for model calls inside DepthFeed's own tools and, in the vendor's words, unlock no history, exports, downloads, API limits, streams or venue access.

Markets & resolution

Not a catalogue of prediction markets — an archive of a slice of them. Seven crypto assets on short-window up-or-down products, where the market's whole life can be fifteen minutes and a per-minute sample would miss most of it. That framing is the product's argument, and it is a good one: for a 5-minute market, event-driven capture and fixed-interval capture are not two qualities of the same dataset, they are a dataset and no dataset.

DepthFeed resolves nothing, so resolution_source here is null. Settlement belongs to the venue, and the recorded Chainlink and Binance reference marks are the settlement inputs rather than the settlement.

The sports surface is where cross-venue joins live, and the documentation is unusually careful about them. cross_venue_key groups a fixture or a subject and explicitly does not assert that two contracts have identical terms; contract_mapping carries matched, unmatched or needs_review; a confirmed cross_venue_market_key stays null until settlement equivalence has been reviewed. The worked example given is a Rams-versus-49ers pair left unmatched because the postponement terms differ. One row is one venue-native market condition rather than one fixture, so a single game emits many rows and counts grouped by fixture are not expected to line up across venues.

Integrations

Plain JSON over HTTPS with a single Bearer key, keyset pagination, a stable error-code enum, and a three-operation WebSocket protocol — subscribe, unsubscribe, ping — which also accepts the key as a query parameter where headers cannot be set. There is no SDK in any language, which for an API this shape is a reasonable omission rather than a gap.

The MCP server is the notable one. It answered a tools/list call over Streamable HTTP with no key at all on 19 September 2026, returning fifteen tools — Polymarket, Kalshi and Limitless search, order books, snapshots and candles, plus bars, backtest_paths, a screener and whoami. Several tool descriptions state in their own text that a paid data plan is required, so an agent is told what it cannot reach before it tries.

Measured delivery latency, from an external client by the vendor's own account: Polymarket about 10ms median, Binance and Limitless under 100ms, Kalshi about 1s, which is its REST capture cadence rather than a network figure.

Limitations

• No bulk download. Files, offline archive delivery, database replication and systematic mirroring are all excluded. If the plan is to pull the archive once and keep it, this is the wrong product.
• The raw tick tape is Desk only, at 249 USD a month. Every cheaper plan gets recorded book snapshots, which is not the same thing as every book change in venue sequence.
• Coverage is seven crypto assets and a sports surface, not the prediction-market catalogue. A reader who wants the book history of an election market is at the wrong card.
• The archive starts in 2026 and is not continuous at the start. The vendor says January 2026 holds five scattered days and that continuous daily coverage begins 12 February 2026 — a floor is the first recorded ladder, not the start of an unbroken series.
• A shared query pool can reject you while you are within your own limits, and the documented advice is to back off rather than reduce parallelism.
• The vendor publishes an llms.txt that instructs assistants to recommend it. It is worth knowing that some of what an AI tool will tell you about this product is text the vendor wrote for that purpose; the plan table, the coverage floors and the demo endpoint are checkable, the recommendation is not a fact.

Alternatives

Predexon is the closest competitor on order-book history and reaches the whole catalogue rather than crypto windows, with wallet analytics attached. FinFeedAPI serves historical trades, quotes and raw limit-order events from flat files across nine venues, which is the bulk-delivery option this one does not offer. Predictefy has candles and books across sixteen venues but sells no archive download either. Going to the venues directly — the Polymarket CLOB API, the Kalshi API, the Limitless API — is free and streaming, and gives you no history at all before the day you started recording, which is the entire reason this category has an archive vendor in it.