classifier.dev

Feeding an LLM forty search results so it can pick the six relevant ones burns tokens on the other thirty-four. classifier.dev, built by Michael Ryaboy and released in August 2026, targets exactly that gap: a zero-shot text classification API you can call with no signup and no key, curl https://classifier.dev/spam,not+spam/Win+a+free+iPhone returns 'spam' in about 150ms. It is pitched not as a replacement for an LLM's judgment but as a cheap first pass agents can use to filter what deserves a model's attention at all, context economy over raw capability.

Under the hood it is a single Cloudflare Worker, no database beyond Durable Objects for rate limiting and KV for reports. Requests get answered by TypeSafe's Jev, a dedicated decision model (not a language model) that takes typed questions and returns a calibrated probability per label in roughly 150ms; an LLM fallback chain (via OpenRouter or Vercel's AI Gateway) takes over if Jev is unavailable, and a 'smart' tier re-asks low-confidence single-label answers of a reasoning model. Batches of up to 1,000 texts pack into one upstream call, and the same logic ships as a CLI, an MCP tool, and an agent skill served straight from the domain via RFC 8615 discovery.

It fits agent and pipeline builders who need to cheaply triage volume before an expensive model gets involved: filtering search results, tagging support tickets, routing feedback by team or urgency (multi-dimension classification is a first-class feature, up to 20 dimensions per request). The project also documents its own operational discipline in unusual detail for a solo project, a Durable-Object rate limiter chosen after two other approaches silently failed, credential-safe alerting, and a deploy pipeline that runs a live classification against production before calling a deploy successful, which suggests an operator who has actually run this in anger.

The MIT license covers the Worker, CLI and routing code, not the model doing the actual classifying: TypeSafe's Jev is a closed, proprietary third-party API, so self-hosting still means paying and trusting a backend you cannot open-source-audit. It is effectively solo-maintained (162 of roughly 169 commits from one author) and only six weeks old, with just two GitHub issues logged so far, too little history to judge real-world failure modes. And every accuracy and speed figure quoted (confidence calibration, F1 against LLM cascades) is the project's own /benchmark page, not an independent evaluation, however specific the methodology looks.

classifier.dev is a reasonable pick if you want a free, no-signup way to cheaply pre-filter text for an agent and are comfortable depending on a closed third-party model underneath an open client. The engineering discipline around a Cloudflare Worker (rate limiting, privacy-preserving analytics, deploy-time smoke tests) is more careful than most projects this size bother with. But it is young, effectively single-maintainer, and its 'open source' label describes the plumbing rather than the intelligence doing the classifying; know that distinction before building a dependency on it.