# Open-Source Prediction Market Tools Cards whose own source is published under a licence you can read — and why that answers a different question at a client than it does at a platform. *https://predictionmarkets.tools/collections/open-source* ## What this collection selects `access.open_source` is a claim about **the thing on the card**: its own source is published under a licence anyone can read. That is all it asserts. It is not a claim that the product is free, that it is maintained, that it can be self-hosted, or that anyone has audited it. Three facts hide behind the phrase, and they come apart constantly. ## A licence, an open repository, and which part is open **The licence says what you may do.** Most of this listing is MIT. [Confido](https://predictionmarkets.tools/tools/confido) is AGPL-3.0, and that is the one worth reading before you plan anything: it reaches code run as a network service for other people, which is precisely what a self-hosted forecasting workspace is. [Metaculus](https://predictionmarkets.tools/tools/metaculus) is BSD-2-Clause, [the alpha bot](https://predictionmarkets.tools/tools/polymarket-alpha-bot) is Apache-2.0. **An open repository says whether you can participate.** The two are independent, and the cleanest demonstration in the catalogue is a card that is *not* in this collection: [kalshi-python](https://predictionmarkets.tools/tools/kalshi-python) is the exchange's own package, its wheel declares its licence as `LicenseRef-Proprietary`, and the repository its metadata names returns 404 to anyone outside the company. You can install it and read the generated Python inside the wheel. You cannot see its history, file an issue against it, or send a patch. Readable code, closed project. **And there is the question of which part is open.** [FinFeedAPI](https://predictionmarkets.tools/tools/finfeedapi) publishes MIT-licensed generated SDKs covering a dozen languages, while the data and the service behind them are paid and hosted — the SDK being open source does not make the service open source. [Polyrama](https://predictionmarkets.tools/tools/polyrama) is the same shape from the other end: its MCP server is MIT, its terminal and API are not. Neither card is in this collection, because the flag is about the product the card is for. ## Open source at a client is not open source at a platform Both kinds are in this listing and they answer different questions. **At a client, it is a question about your machine.** [py-clob-client](https://predictionmarkets.tools/tools/py-clob-client), [polymarket-client](https://predictionmarkets.tools/tools/polymarket-client), [pykalshi](https://predictionmarkets.tools/tools/pykalshi), [the Limitless SDK](https://predictionmarkets.tools/tools/limitless-sdk), [CCXT](https://predictionmarkets.tools/tools/ccxt) and [PMXT](https://predictionmarkets.tools/tools/pmxt) all run in your process, holding your credential. Open source means you can answer three things yourself: what the code does with the key, what it pulls in alongside it, and whether you can fork it when the vendor stops. That third one is not hypothetical here — see the archived card below. **At a platform, it is a question about the house rules.** [Manifold](https://predictionmarkets.tools/tools/manifold) publishes the code its own site runs on, which is why its fee position is a fact rather than a promise: the taker fee constant and the flat trade fee in that code are both zero. [The Manifold API](https://predictionmarkets.tools/tools/manifold-api) and Metaculus are the same situation. What none of them gives you is the thing that matters most about a platform — the balances, the users, the questions and the resolutions all live on a server you do not run. Metaculus makes the limit vivid: since 2026 every API request needs an account token, and the licence on the source had nothing to say about that. The sharpest case is [brier-fyi](https://predictionmarkets.tools/tools/brier-fyi), which scores four venues against each other. It is MIT, and the code is why anyone can state what is wrong with it: the last commit is dated 24 January 2026 and the one before it is titled "infra, pause nightly data pipeline". The site is up and the numbers are not moving. A closed product in that state would simply look stale. ## What to check, in the order it goes wrong - **Is it archived, and does the README say anything?** [py-clob-client](https://predictionmarkets.tools/tools/py-clob-client) is MIT, archived, and carries Polymarket's own notice that it no longer works and should not be used. This is the official SDK for the largest venue in the catalogue, and its licence is unchanged and irrelevant. The replacement is [polymarket-client](https://predictionmarkets.tools/tools/polymarket-client). - **Do the releases and the commits agree?** [polymarket-us-python](https://predictionmarkets.tools/tools/polymarket-us-python) has exactly two PyPI releases, 0.1.1 and 0.1.2, both uploaded on 22 January 2026, with no tags or GitHub releases since, while `main` took a retry and correlation-id change in June 2026 and an order-snapshot subscription fix in September — so what you read on GitHub is not what `pip install` gives you, and the card's advice is to install from git if you need the fixes. [The Limitless SDK](https://predictionmarkets.tools/tools/limitless-sdk) is the opposite and the standard to hold the rest to: version 1.1.1 on 17 September 2026, which is also the date of its last commit. - **Does an open issue tracker actually tell you anything?** [PMXT](https://predictionmarkets.tools/tools/pmxt) shipped 322 releases between January and July 2026 and then stopped; it is not archived and its tracker is open, with 1,236 open issues and 77 open pull requests counted on 19 September 2026. An open tracker that large cannot tell you whether your problem is already known, which is the only thing anybody opens a tracker for. ## What happens to your keys in code you run yourself This is the sentence the collection exists to print, and it is the same one the [clients and bots](https://predictionmarkets.tools/categories/trading-clients-and-bots) category page makes: **an MIT licence is permission to read the code, not evidence that anyone did.** What that means in practice, from cards in this listing: - **Prefer a library that separates reading from trading.** `polymarket-client` splits it in the type system: its public clients need no credentials at all, while `SecureClient` and `AsyncSecureClient` take a local private key and can trade. A notebook pulling prices never touches key material. - **Read where the credential is asked for.** PMXT's quickstart puts a `pmxt_api_key` and a `private_key` into the same constructor; the documentation says PMXT's servers never see the private key and does not explain the signing path that would make that true. Self-hosted mode exists for exactly this, and it is in the same package. - **Count the dependencies.** They run in the same process as the key. `pykalshi` has five at runtime — `httpx`, `pydantic`, `cryptography`, `python-dotenv` and `websockets` — which is a list a person can actually check. - **Know which credential you are handing over.** An exchange API key that can be scoped to trading is a different exposure from a wallet private key that can also move funds. On the on-chain venues, the key that signs an order is the key that holds the collateral. Nothing in this collection has been run against a funded account by this site. Licences, release dates, commit dates, archive status and dependency lists are read from the repositories and the published packages, and each is dated on the card it came from. ## Cards - [Brier.fyi](https://predictionmarkets.tools/tools/brier-fyi.md) — Brier scores and letter grades for matched questions across four platforms. - [CCXT](https://predictionmarkets.tools/tools/ccxt.md) — One client for seven prediction venues, inside a 104-exchange crypto library. - [Confido](https://predictionmarkets.tools/tools/confido.md) — Open-source forecasting workspace you host yourself, with a Brier score pointing upwards. - [Fatebook](https://predictionmarkets.tools/tools/fatebook.md) — Write down what you think will happen, in Slack or a browser, and get scored on it. - [limitless-sdk](https://predictionmarkets.tools/tools/limitless-sdk.md) — Limitless Exchange's own async Python SDK - CLOB and NegRisk orders, WebSocket, MIT. - [Manifold API](https://predictionmarkets.tools/tools/manifold-api.md) — Every read is keyless, the server is MIT-licensed, and the currency is play money. - [Manifold](https://predictionmarkets.tools/tools/manifold.md) — Anyone can open a question, anyone can take a side, and the currency buys nothing. - [Metaculus](https://predictionmarkets.tools/tools/metaculus.md) — Proper scoring and public track records on questions nobody can take a position in. - [PMXT](https://predictionmarkets.tools/tools/pmxt.md) — CCXT-shaped client for prediction markets, with a hosted API and a self-hosted mode. - [Alphapoly (polymarket-alpha-bot)](https://predictionmarkets.tools/tools/polymarket-alpha-bot.md) — Self-hosted scanner for covering portfolios across correlated Polymarket contracts. - [Polymarket CLI](https://predictionmarkets.tools/tools/polymarket-cli.md) — Polymarket's Rust command line - MIT in the manifest, with no LICENSE file in the repo. - [polymarket-client](https://predictionmarkets.tools/tools/polymarket-client.md) — Polymarket's own unified Python SDK - sync and async, data through order signing. - [Polymarket US Python SDK](https://predictionmarkets.tools/tools/polymarket-us-python.md) — Official SDK for Polymarket US - installs as polymarket-us, unreleased since January. - [Polymarket US TypeScript SDK](https://predictionmarkets.tools/tools/polymarket-us-typescript.md) — Official TypeScript client for Polymarket US - npm has shipped nothing since January. - [Prediction Market Analysis](https://predictionmarkets.tools/tools/prediction-market-analysis.md) — Open-source Kalshi and Polymarket dataset, indexers and analysis scripts. - [py-clob-client](https://predictionmarkets.tools/tools/py-clob-client.md) — Polymarket's own Python CLOB client - archived, and declared non-functional by its README. - [pykalshi](https://predictionmarkets.tools/tools/pykalshi.md) — Unofficial Kalshi client with what the generated SDK leaves out - streams and retries. - [Quantified Intuitions](https://predictionmarkets.tools/tools/quantified-intuitions.md) — Forecast questions that already resolved, and learn in ninety seconds how wrong you were. ## FAQ ### Does open source mean the tool is maintained? No, and this listing contains the counterexample. py-clob-client is MIT-licensed, and it is also archived, with a notice at the top of its README from Polymarket saying the client no longer works and should not be used. A licence is a permission granted once; maintenance is three dates - the last release, the last commit, and whether the repository is archived - and every card here carries them. ### Is open source safe to hand an API key or a wallet key? Open source is permission to read the code, not evidence that anybody did. A client that signs orders holds the credential in the same process as every dependency it pulls in. Prefer a library that separates the read path from the trade path the way polymarket-client does, prefer exchange keys scoped to trading over wallet keys that can also move funds, and read what the code does with the key before you fund the account. ### What does it mean that a venue is open source? Something narrower than it sounds. Manifold publishes the code its platform runs on, which is where its fee constants can be read rather than taken on trust, and Metaculus publishes its own under a BSD-2-Clause licence. Neither gives you the balances, the users or the resolutions, all of which live on their servers - and Metaculus started requiring an account token on every API request in 2026 without its licence changing at all. ### Which licences are in this collection? Mostly MIT, with Apache-2.0 on polymarket-alpha-bot, BSD-2-Clause on Metaculus and AGPL-3.0 on Confido. The AGPL is the one that changes your plans, because it reaches code you run as a network service for other people, which is exactly what a self-hosted forecasting workspace is. For a library you import into your own program, MIT and Apache-2.0 differ mainly in their patent and attribution terms.