Adjacent vs DepthFeed vs Predexon: what a cross-venue data API sells that the venues give away

Start from what the venues already give away

Every venue in this catalogue publishes its own API and most of it is free and keyless. Polymarket's CLOB answers book, price, midpoint, spread and last-trade calls unauthenticated; its Gamma catalogue needs no key at all; Limitless serves markets, books and history without headers and runs open public WebSocket channels; Kalshi opens its REST market data and gates only the socket and the orders. Full depth, live, for nothing.

So the question is not which aggregator is best. It is what any of them sells that is not already free, and there are exactly three honest answers: a schema in which two venues can be compared, history from before the day you started recording, and one integration instead of five. The three services above answer those three questions very differently, and one of the three answers is worth less than it sounds.

Normalisation, and the version of it that is honest

All three hand you one schema. What separates them is what they claim about a row.

Adjacent normalises two venues and prefixes every id with the venue it came from, so a list response feeds a detail or history call without rebuilding a ticker, and prices arrive on a 0-100 scale. On top of that sits the part nobody else here publishes: 64 indices, cross-venue reference rates, and a stored news and filings corpus linked to the markets each item touches. That is construction rather than normalisation, and it comes with a methodology you are adopting — fixed composite weights, a factor set inside the sub-indices, and a scale on which the number is not a probability until you subtract 50. Its own caution is the right one: check an index's constituents before quoting its level at anyone.

Predexon normalises five venues, and the coverage behind that number is lopsided. Markets, order-book history and trades exist for all five; the WebSocket, the oracle feed, the wallet and smart-money analytics and the builder endpoints are Polymarket-only, and the tick archive covers three of the five. "Five venues" is true of the catalogue endpoint and false of almost everything else on the service.

DepthFeed does the least normalising and is the most careful about it, which is the interesting combination. On its sports surface a cross-venue key groups a fixture and explicitly does not assert that two contracts have identical terms; a mapping field carries matched, unmatched or needs-review; and a confirmed cross-venue market key stays null until settlement equivalence has actually been reviewed. Its worked example is a pair of contracts on the same game left unmatched because the postponement terms differ.

That is the whole argument about normalisation in one example. Two contracts on the same event with different postponement terms are not one row, and a service that returns them as one row has not normalised anything — it has hidden the difference you were paying it to surface. Adjacent does the same job in another shape: a reference rate returns the spread between its sources, and its docs say a wide one means the number is hiding a disagreement rather than summarising one. Predexon makes no such claim because it does not match markets across venues at all; what you get is five venues in one shape, with the joining left to you.

History, which is the only thing here you cannot rebuild

The venue APIs are free and streaming and give you nothing before the day you started recording. Limitless states the sharpest version of that: its price history is clipped to each market's own lifespan, so a 15-minute crypto market returns about 15 minutes, and any longer series has to be stitched together across the separate markets that made it up. This is the reason this category has archive vendors in it.

DepthFeed and Predexon are both archives and they are not substitutes.

DepthFeed records the complete ladder, both sides and every level, at every change rather than sampled on a clock, across crypto up-or-down windows on three venues and seven assets. Its argument is that for a 5-minute market, event-driven capture and fixed-interval capture are not two qualities of one dataset but a dataset and no dataset, and that is correct. Two response mechanics decide whether a backtest built on it is honest, and both are on the card: an interval parameter that selects rather than interpolates, so a bucket appears only if a book was captured in it; and a forward-fill mode that produces a dense grid instead and stamps each row with how stale it is. Every interval response reports buckets in range, buckets filled and whether the window is complete. The floors matter as much as the depth — the longer series begin 2026-01-03, but the vendor states January holds five scattered days and continuous daily coverage starts 2026-02-12. A floor is the first recorded ladder, not the start of an unbroken series.

Predexon sells the same substance in the opposite delivery: every placement, update, cancel and fill as raw deltas in compressed Parquet, pulled in bulk rather than paged over HTTP. Its floors are later and its venues are wider — Polymarket from 2026-03-02, Kalshi from 2026-03-05 with lossless sub-cent data from 2026-04-20, Opinion from 2026-03-10, and Kalshi book snapshots only from 2026-09-18. Days are finalised the following morning, so today is a 404 rather than a partial. And it carries the one series nothing else in this catalogue publishes: resolution as data. Oracle proposals, disputes, settlements and resets stream as events, and questions can be filtered by oracle state. If your exposure is to the resolution rather than to the price — a disputed question is four to six days of locked capital — that feed, not the book, is the reason to be here.

Adjacent is barely in this contest and does not pretend to be. Free price history is hourly buckets for the previous 30 days or daily for the previous 90; finer intervals and anything deeper are paid. That is a chart, not an archive.

The third reason is the weakest, and it should be said plainly

"It does not break when the venue changes an endpoint" is the reason most often given for paying an aggregator, and none of these three sells it.

Predexon has already retired its v1 behind a 410 with a pointer to v2. Its card reads that as a sign of a maintained API, and it is — but a migration is a migration whoever hands it to you, and this one came from the vendor rather than from the venue. DepthFeed is on its third version. Adjacent publishes no terms of service at all: on 2026-09-19 its terms, legal and privacy paths each answered 200 with the landing page, byte for byte the document a nonsense path is served, so there is nothing setting out uptime, notice of breaking changes, or what happens to your access if the vendor's own upstream arrangement with a venue changes.

What you actually buy is one integration instead of three or five, and somebody else on call when a venue moves. That is worth real money. It is not stability, and the service with the most visible version discipline is the one that has already made you migrate once.

The licence is the part of this that bites hardest, because it is the part a reader assumes they bought. Adjacent states on its pricing page that neither paid plan permits redistribution or use in a financial product, which rules out most of what a data licence is normally purchased for and leaves research, internal tooling and agents. DepthFeed describes itself as an independent research project with no affiliation to or endorsement from any venue it records — which is a licensing fact before it is a modest one, because what a venue's own terms permit you to do with its prices is not settled by a subscription to somebody who recorded them.

The free tiers in one unit, and what runs out first

Converted to requests a minute, the free tiers are closer than the prices suggest: Adjacent allows 50 requests per 10 seconds per IP, which is 300 a minute; DepthFeed's Explorer and Predexon's Free plan both allow 1 a second, which is 60. Neither DepthFeed nor Adjacent states a daily quota on the free tier. Predexon does: 1,000 premium requests a month, which is about 33 a day.

On none of the three is the request rate what runs out first.

On Adjacent it is coverage. On 2026-09-19 every one of the 1,694 markets the public endpoints returned was Kalshi, and a request filtered to Polymarket came back empty. The free tier does not demonstrate the coverage the paid tier is bought for, and 300 requests a minute of one venue is not two venues. The 15-minute delay is the second wall and the shallow history the third.

On DepthFeed it is the snapshot floor. Explorer caps the snapshot interval at 60 seconds, and the product is markets whose whole life can be five minutes. That is about five observations of an entire market — the free tier structurally cannot see the thing it is a sample of. Seven days of history and a single live subscription run out at roughly the same moment.

On Predexon it is the monthly budget, which is oddly shaped. A thousand premium requests a month is one polling loop spent before lunch. But every list-market and order-book-history endpoint across every venue is unmetered on every plan including this one, so the expensive thing to serve is the free thing. The free tier will not support a dashboard and will support a backfill, which is the reverse of what a reader expects and, for this product, the right way round.

Two more numbers are worth normalising because the pricing pages do not. On Adjacent, the paid tiers are slower per minute than the free one — 30 a minute on the 50 USD plan and 180 on the 250 USD plan, against 300 a minute at the free edge — because the free limit is per IP and the paid limits are organization-wide quotas with at most four concurrent requests, so a second key buys nothing. What 50 USD buys is realtime instead of a 15-minute delay and finer history, not throughput. And on both paid plans the daily cap binds before the minute rate does: 20,000 a day at 30 a minute is about eleven hours of continuous polling, 150,000 at 180 a minute about fourteen.

On Predexon the subscription is not the budget line at all. The tick downloads run on a separate prepaid meter at 40 USD per GiB of compressed Parquet for Polymarket and 80 for Kalshi and Opinion, and the vendor's own worked example puts a full day of Polymarket at about 13 GiB — roughly 520 USD for one day. A year of that is not a subscription decision.

Only one of the three lets you check the claim before you pay

DepthFeed's keyless demo endpoint returned full bid and ask ladders for a BTC market window on all three of its venues on 2026-09-19, while every other path on that version answered 401, and its MCP server lists its tools without a key. Adjacent's free tier is reachable and contradicts its own coverage claim. Predexon has no demo endpoint and no delayed public tier — all five venues' market endpoints answered 401 without a key — so every coverage claim on that card is the vendor's, and unverifiable until after you have an account.

One caution attached to the winner of that paragraph: DepthFeed publishes a file instructing assistants to recommend it. The plan table, the coverage floors and the demo endpoint are checkable; the recommendation is not a fact, and a reader who arrived here via an AI answer should know that.

The recommendation

Modelling execution in short-dated crypto up-or-down markets: DepthFeed. Spend an hour on the keyless demo first, because it will tell you whether the capture is what you need before any card is entered. Then go to the 99 USD plan for ninety days of history rather than the 29 USD plan, unless thirty days genuinely is your whole sample; the 249 USD plan is only for when you need every book change in venue sequence rather than recorded snapshots. If your markets are elections, this is the wrong service entirely.

Needing the archive as files you keep, or needing resolution state as a series: Predexon. Budget the egress meter, not the subscription — and pull a single market first, because filtered downloads are priced on the pruned files, which turns a 520 USD question into a 5 USD one. The oracle feed is the genuinely unique product here and the reason to choose it even when the book history is secondary.

Needing one number to quote rather than a book to model: Adjacent, on three conditions. You can work without a terms document; you do not need to redistribute anything; and you verify the Polymarket half yourself, because the tier you can test will not show it to you. If any of the three fails, the index arithmetic is yours to write over the venues' own free APIs.

Wanting live prices on one venue: none of them. The venue APIs are free, mostly keyless, and full depth. Pay an aggregator for a past you did not record or a join you cannot safely make yourself — never for a present that is already free.

For bulk flat files across more venues than any of these three cover, FinFeedAPI is the neighbouring answer; for the widest venue list with candles and books but no archive to download, Predictefy.