What to use instead of Metaforecast

The product

Metaforecast was a search box. You typed "Trump" or "semiconductors" and got back probabilities on that subject from everywhere at once — the venues and, unusually, the forecasting platforms alongside them. The fetchers in its own repository name the spread: Metaculus, Manifold, Good Judgment and Good Judgment Open, INFER, PredictIt, Kalshi, Polymarket, Smarkets, Betfair, Insight Prediction, Rootclaim, Guesstimate, FantasySCOTUS and a handful of individual forecasters' question sets. Two of those now have pages of their own here, because they are gone too: INFER and Insight Prediction.

It was built by the Quantified Uncertainty Research Institute, released under the MIT licence, and it gave away everything: a public GraphQL API, dashboards other sites embedded, a Discord bot integration, a Twitter account that posted matches. Its README describes a database refreshed once a day at 3am UTC. It also carried star ratings — the README is explicit that these were the maintainer's judgement of how much a given aggregate deserved to be trusted, not a computed score — which is the sort of thing only a non-commercial project ships.

How it ended, and why the date is a date and not a day. On 20 June 2026 the institute published a post saying it was moving into maintenance mode after about seven years: Guesstimate, Squiggle Hub and RoastMyPost would keep running, and "Foretold.io and Metaforecast will be wound down. These haven't been functioning for some time; now we'll take steps to formally end them."

That sentence is the whole obituary, and it means the service stopped working before anybody said so. The frontmatter on this page records 20 June 2026 because it is the only date anybody published — not because the search box answered on 19 June. What can be dated around it: the standalone repository is archived on GitHub with its last push on 9 February 2025, and on 21 September 2026 metaforecast.org answered with a 308 redirect to the institute's blog, taking the GraphQL endpoint with it. The real end is somewhere in that window, and nobody recorded where.

What to use instead

The first thing to accept is that the catalogue cannot give you back the half of Metaforecast that made it interesting. Commercial feeds aggregate venues, because venues have volume and volume is what somebody will pay for. Metaculus, Good Judgment Open and the rest have no order book to sell, so almost nothing indexes them.

If what you want is one question across several platforms, Brier.fyi is the closest thing left, and it is the only pick that reaches past the venues: it hand-matches questions across Kalshi, Polymarket, Manifold and Metaculus and scores each platform's accuracy on them. It is MIT and self-hostable, which is the same bargain Metaforecast offered. Read its card before you rely on it, because its pipeline is paused — this is a rescue, not a like-for-like.

If what you want is breadth across venues and you will pay for it, Predictefy puts sixteen venues behind one REST, WebSocket and MCP layer, and FinFeedAPI normalises nine into one REST schema with order books, OHLCV and historical flat files. Both are the opposite of what you had in one specific way: nothing reads without a key, and the bill scales with how curious you are.

If what you want is the property that mattered — that it was yours — take the open-source route and do not rent the layer again. PMXT is MIT-licensed across fifteen venues in Python and TypeScript with a self-hosted mode, and CCXT carries a prediction namespace over seven event venues inside a library a lot of people already have installed, in eight languages. Neither searches questions the way Metaforecast did; both mean the next maintenance-mode announcement is not your problem.

If you only ever used it for the play-money platforms, go direct. Manifold's API reads without a key at all and its server is MIT, which is why Metaforecast could index it in the first place. Metaculus has an API too, and its card records that it now requires a token.

If what you wanted was one number rather than a list, Adjacent publishes indices and reference rates across venues, with Kalshi on a tier you can test before paying. That is a different object from a star-rated aggregate — it is a computed index with a published method, not somebody's judgement of how much to trust a crowd — and for some uses it is the better one.

What you give up

Coverage of the forecasting platforms. This is the loss and everything else is detail. A search that spanned Metaculus, Good Judgment Open and INFER next to Polymarket let you see a question priced by people with money and estimated by people with a track record, side by side, and notice when they disagreed. Of the six picks, exactly one reaches two of those platforms, and its pipeline is paused. If that comparison was your workflow, it is not available at any price here.

Free, keyless and open at the same time. Metaforecast asked for nothing: no account, no key, no card, and the source was MIT if you disliked how it worked. The picks split that bargain up. Brier.fyi, CCXT, PMXT and the Manifold API each keep part of it; Predictefy and FinFeedAPI keep none of it and are better funded for exactly that reason. Whichever you take, you are choosing which half to give up.

Somebody's opinion, stated as such. The star ratings were a person telling you which aggregates deserved weight, in public, with their name on it. Every replacement here reports numbers and declines to say which ones to believe. That is more defensible and less useful, and nothing in this catalogue is going to hand you back a curator.

The embeds and the bots. Metaforecast was designed to be put inside other things — dashboards on somebody else's site, a Discord command, a forum post. The APIs here are designed to be called by code you write. If what you had was a !metaforecast command in a server full of forecasters, what you now have is a small project.

Migration notes

Nothing needs to be withdrawn, and that is the only easy part. There was no account, no balance and no data of yours inside it. What breaks is code and links: anything pointing at the GraphQL endpoint now follows a redirect to a blog and fails in whatever way your HTTP client fails when it gets HTML instead of JSON. Find those call sites before they find you.

Do not expect an archive. The institute's post names no data dump and no snapshot, and the question histories the site rendered are not published anywhere that we could find. If you were using it as a record of how a probability moved, that record is not recoverable from the project — it has to be rebuilt from each platform's own API, where the history exists at all.

If you fork it, fork from the monorepo. The archived standalone repository is explicit that the code moved, and it was last pushed in February 2025. The live copy sits in QURI's Squiggle monorepo under apps/metaforecast. Either way you are adopting fetchers written against platform APIs as they stood more than a year ago, several of which have changed; budget for that first rather than for the hosting.

The institute asked. The same post says that anyone interested in improving, reviving or maintaining these projects should get in touch. That is an unusual sentence to find in a shutdown notice and it is still the shortest route to the thing actually existing again.