PolyOrderbooks

What it is

An archive of Polymarket's order book for one slice of the venue: crypto markets, captured off the CLOB WebSocket feed and served back as bucketed history over REST. It is a live product with a live API — api.polyorderbooks.com answered this machine on 20 September 2026, returning {"error":"unauthorized"} with HTTP 401 to a keyless call and "Invalid API key." to a wrong one, which is the pair of answers a running gateway gives.

The read surface is small. GET /v1/series, /v1/events, /v1/markets and /v1/tags discover; GET /v1/markets/{id_or_slug} resolves a market to its outcome token ids; then /markets/{id}/prices, /markets/{id}/metrics and /markets/{id}/books return the history, with /tokens/{id}/prices and /tokens/{id}/books for the single-token form. GET /v1/usage echoes your plan's limits. Every one of them takes start_ts, an exclusive end_ts and a resolution from 1s through 1d, and pages with opaque cursors at a default of 100 rows and a maximum of 200.

Two design statements from the documentation are worth carrying into any code you write against it. Reads are local: discovery and history query the vendor's own store and never call Polymarket live, so a market that was never captured does not appear. And captures are sparse by design, materialised uniform: the collector writes a row when a token's book changes, and the API fills the gaps by last-value carry so that every bucket you asked for comes back. A 1-second series is therefore a grid over an event stream, not a proof that the book was read on each of those seconds.

Beside the API sits Backtest AI, a natural-language strategy replay: you describe a crypto up-or-down rule in a sentence, it compiles that into a specification, replays it over resolved markets against the recorded ladders, and returns net P&L, win rate, profit factor, maximum drawdown, an equity curve and a written summary. The response carries the compiled strategy block, which is the part that makes a run checkable rather than a number to take on trust.

Availability

No geographic gate and no identity check. The terms of service dated 23 August 2026 condition eligibility only on being able to form a binding contract and meeting the age requirement of your own jurisdiction; there is no US clause, no restricted-country list and no KYC step. Access is an email-verified account and a key beginning pob_. Paid plans take cards or, per the vendor, 300-odd cryptocurrencies.

What there is instead of a geographic gate is silence about who you are contracting with. The terms name no company, no country and no address: governing law is "the laws applicable where PolyOrderbooks principally operates". They also still describe paid plans as renewing "when checkout is live", a month after the pricing page began offering checkout. Read section 15 before an invoice matters to you.

Nothing here is permission to trade anything. Reading a recorded Polymarket book is not access to Polymarket — that question is answered on polymarket, and the terms are explicit that your licence covers the delivery of the service and not ownership of the underlying exchange data.

Pricing

One free plan and one paid axis. Starter is $0, needs no card, and is not throttled on resolution — it can request 1-second buckets exactly like every paid plan. It gets three days of lookback, 60 requests a minute and 1,000 a day.

Paid plans are the same product with a longer rolling window: $19 a month for 30 days, $29 for 60, $39 for 90, $49 for 120, all at 300 requests a minute and 50,000 a day. Two speed packages sit on top — Fast at $19 a month for 1,000 a minute and 120,000 a day, Heavy at $39 for 2,500 and 300,000 — and those numbers are in the documentation rather than on the pricing page, which advertises the add-ons without saying what they buy. Backtest AI is $19 a month as an add-on to a data plan, or $29 standalone with its own 30-day window. Enterprise is a conversation: S3 bulk drops in Parquet, CSV or JSON, custom limits and an SLA.

Two things the price table does not say out loud. The window is a rolling lookback from now, not a slice you choose: a query whose start_ts predates it returns HTTP 403 naming the earliest timestamp allowed. And 120 days is the deepest history anyone can buy without talking to sales, which puts a ceiling on any study that wants a year.

Rate limits are per minute on a rolling window and per day in UTC, resetting at midnight, with 429 and a Retry-After header in seconds.

Markets & resolution

Polymarket crypto markets only. The pricing FAQ and the documentation both say it in the same words: other categories are not in the archive. That is the single fact that decides whether this product is for you — there is no politics book here, no sports, no second venue, and no plan that adds one.

PolyOrderbooks resolves nothing, so resolution_source is null. Settlement belongs to Polymarket and is described on polymarket-clob-api. What the archive keeps after an event settles is the market and its winner, which is what makes a replay over resolved markets possible at all.

The site's own counter reported 819,500 markets indexed, 779.7M order book snapshots, 144.6M price snapshots and 332.8M metrics snapshots on 20 September 2026.

Separately from the API, the vendor publishes a citable dataset rather than a sample: 897,192 one-second L2 snapshots across 805 resolved Polymarket up-or-down markets at three contract lengths — 5 minutes, 15 minutes and 4 hours — over eight coins, captured 21 to 24 August 2026, deposited on Zenodo on 24 August 2026 under CC BY 4.0 with a DataCite DOI, and mirrored as Parquet on Hugging Face and CSV on Kaggle. Its README is unusually candid about the capture: the Polymarket delta feed carries roughly 24,000 events a second with no sequence numbers, so dropped messages cannot be detected from the stream, and the collector re-reads the full book over REST every 60 seconds to bound the drift. It also publishes its own error rates rather than cleaning them — 16.9% of five-minute rows have one side of the book empty, rising to 76.2% in the final minute, and 1.5% to 3.24% of rows are crossed and flagged rather than removed.

Integrations

Plain JSON over HTTPS with X-API-Key: pob_… or the same value as a bearer token; dashboard session tokens are explicitly not accepted on data routes. There is no WebSocket and no live feed of any kind — this is history and only history.

An official Python client is on PyPI as polyorderbooks, MIT by its metadata, version 0.2.0 uploaded 24 August 2026, one dependency (httpx) and Python 3.9 or newer. An official MCP server is on npm as @polyorderbooks/mcp-server, MIT, version 0.2.3 published 25 August 2026, exposing the archive as tools in Claude and Cursor. Backtest AI has its own route, POST /v1/ai/backtest, available on any authenticated plan.

The documentation publishes an OpenAPI description and an llms.txt index, and the endpoint pages carry a request playground once you paste a key.

Limitations

• Crypto only, one venue only. No politics, no sports, no macro, no Kalshi, no Limitless. If the question is anything but a crypto up-or-down market on Polymarket, this product has no answer at any price.
• 120 days is the self-serve ceiling. Longer history exists only as an Enterprise contract, and the free tier's three days is short enough that validating the schema is most of what it is for.
• Nothing is keyless. Every /v1 route requires an API key, so there is no way to see a real response before creating an account. Both keyless and wrong-key calls returned 401 on 20 September 2026, which proves the gateway is up and proves nothing about the data behind it.
• "Snapshotted every second" is a bucket, not a read. Capture is on book change with a 60-second REST reconcile; the uniform 1-second series is produced by last-value carry. The published dataset says so plainly and quantifies the drift; the marketing page does not.
• The advertised source repositories do not exist. The documentation, the llms.txt, the PyPI project links and the npm repository field all point at github.com/polyorderbooks/…, and on 20 September 2026 that account returns 404 — so the "runnable examples on GitHub" and the SDK source are not there to read.
• No WebSocket, no alerts, no live prices. Anything happening now comes from Polymarket itself.
• 200 rows a page. At 1-second resolution that is 200 seconds per request, so a single day of one market's book is 432 calls — comfortable inside the paid daily cap, and a third of the free plan's entire day.
• A counterparty you cannot name. No company, no jurisdiction and no address anywhere in the terms, on a product whose value is that its archive keeps existing.

Alternatives

DepthFeed is the closest thing to a direct competitor and differs on three axes that matter: it covers Polymarket, Kalshi and Limitless rather than one venue, it serves the raw tick tape of every book change on its top plan rather than materialised buckets, and it has a keyless demo endpoint you can call before signing up. It costs more — $29, $99 and $249 against $19 to $49 — and it does not publish a dataset with a DOI.

For live books rather than recorded ones, the Polymarket CLOB API is keyless, free and authoritative, and keeps no history worth the name; the Gamma API is the catalogue beside it.

For breadth rather than depth — many venues, many subjects, no ladders — Adjacent and FinFeedAPI sell a normalised cross-venue feed, and neither is free.

If what you actually need is a citable file rather than a subscription, take the Zenodo deposit and stop: it is CC BY 4.0, 43 MB, carries a persistent identifier and needs no account.