CCXT gives 104 exchange integrations one common vocabulary
CCXT's README lists 104 cryptocurrency exchange integrations and a separate set of 9 prediction-market APIs. Applications get common methods for tickers, order books, trades, candles, balances, and orders instead of maintaining a different client for every venue. That is useful: a team can add another market without redesigning its whole connector layer. CCXT remains library software, so it does not supply a trading strategy, portfolio rules, storage, or production monitoring.
Version 4.5.85 covers JavaScript, TypeScript, Python, C#, PHP, Go, Java, and Rust. Several implementations are generated from the TypeScript source, which explains part of the repository's size. Public REST calls work without an account. Private methods need exchange-issued credentials, and some exchanges require identity verification before providing them. WebSocket classes cover streamed data, while a CLI and local MCP server provide other ways to call the same exchange layer.
What happened when we ran it
Our sandbox installed commit 70b73bf in 25 seconds. Npm added 498 packages and the installed tree occupied 338 MB. The environment was a fresh, unprivileged Debian container with 3 CPUs, 8 GB of RAM, Node 22, and no secrets. Npm audit found 0 known vulnerabilities across critical, high, moderate, and low severity levels. Installation was the only main step that completed successfully.
The build failed after 45 seconds, and the test command failed after 39 seconds. Both paths reached ESLint 9.39.4 and stopped with SyntaxError: Invalid regular expression flags while loading an ES module. The supplied log tail does not name the expression, file, or underlying cause, so blaming Node, ESLint, or project code would be a guess. The test path never produced a completed test summary because it stopped at the same lint error.
The 439.3 MB source tree is contributor-sized
The checkout contained 9,343 files and about 6,751,569 lines of source before installation. It occupied 439.3 MB, then the npm dependencies added another 338 MB. Our scan found 16 CI workflow files, a Dockerfile, and a compose file, but no tests directory. That last signal describes the directory scan, not an absence of test code elsewhere in a repository whose build scripts and generated language targets span several layouts.
Package users have a much shorter path than contributors. The README shows one-command installs from npm, PyPI, Packagist, NuGet, Go modules, Maven, and crates.io. Docker is presented as a development option for the cloned repository, not a requirement for applications importing CCXT. Our 25-second npm install supports that consumer story. It does not rescue the source build, which failed before completion on the same commit.
One API still needs venue-level data checks
CCXT smooths over naming, authentication, and response shapes, but exchange differences remain observable in the common fields. Open issue #30673 on OHLCV volume units says candle volume may represent base currency, quote currency, or contracts depending on the exchange and market type. The report says that unit is not available as machine-readable feature metadata. A plausible number can therefore be wrong by 1,000 times after an incorrect conversion.
Streaming deserves the same caution. Open issue #30220 on buffered WebSocket updates reproduces a second message being dropped when 2 updates with the same message hash arrive in one event-loop turn. That report concerns CCXT Pro 4.5.77 rather than our sandbox build, so it is issue evidence, not our measurement. For balances, fills, or positions, write regression tests around the exact exchange, market type, and CCXT version you deploy.
Private calls add credentials, permissions, and fee choices
Public price data needs no key, while private APIs can place orders, transfer funds, and request withdrawals. The official MCP server starts with public data and keeps trading, withdrawals, and raw endpoints behind separate opt-in capability tiers. Its README describes confirmations, notional caps, and an audit journal. Those controls help, but an application still owns exchange permissions, secret storage, IP restrictions, and the decision to disable withdrawals on a trading key.
The README also names 7 builder-program exchanges and says calls through those integrations add 1 bps unless exchange.options['builderFee'] is set to false. That option belongs in a configuration test because it affects cost. Rate limits need similar attention: CCXT defaults to a leaky-bucket limiter and offers a rolling-window mode whose documented default window is 60 seconds. Neither setting guarantees that every venue will accept the same request pattern.
Version 4.5.85 arrived two days before our run
GitHub showed 44,248 stars and a last push on October 3, 2026. Release v4.5.85 shipped on October 1, one day before commit 70b73bf. The open queue held 240 issues and 472 pull requests when checked, which explains the repository API's combined count of 712. Those figures show a heavily used project with current work, not a quiet library living on old coverage claims.
The same v4.5.85 notes include fixes for WOO order cost, BingX history filters, Coinbase International candle calculations, Binance rate-limit costs, and several other venue details. Frequent repair is both a reason to use CCXT and a reason to pin it. The project absorbs exchange churn for you, yet each update can touch money-facing parsing and order behavior. Review release notes and replay your venue tests before upgrading.
Choose a bot platform when connectors are not the whole job
CCXT is the strongest fit when your application already owns strategy and operations but needs access to more than 100 venues. Freqtrade adds Python strategy development, backtesting, and bot operation. Hummingbot centers bot deployment and market-making workflows. Cryptofeed narrows the problem to normalized WebSocket market data, which is attractive when private trading methods would only widen your security surface.
Start with a published CCXT package, one exchange sandbox, and the smallest API permission set that works. Our run gives contributors a different instruction: do not call commit 70b73bf green on Node 22, because its build and test commands both stopped before completion. The abstraction saves connector work only if your own checks catch the venue-specific mistakes that a common method name cannot remove.

