The classifier.dev CLI. Sort text into your own labels from the shell — up to a thousand fast-tier lines per request (smart batches are capped at 200 for the public quota), calibrated confidence when available, and no API key. Multi-label answers carry a score per label instead.

npm i -g classifier-dev        # installs `classify`
npx classifier-dev --help      # or run it without installing

Use

classify spam,"not spam" "Win a free iPhone now"
spam

classify bug,feature,praise < feedback.txt
bug       0.99    the checkout button does nothing
praise    0.97    love the new dark mode
feature   0.61    would be nice to export as CSV

One line per input, in input order: label, confidence, text, tab-separated (--id adds the id between confidence and text). --json gives NDJSON with everything the API returns; --quiet gives labels only; --count gives a histogram.

Synchronous requests are retained for training and operator review by default, with doubled classification quotas. Use --no-share-data on every invocation to opt out and use standard quotas:

classify spam,"not spam" --no-share-data "A private example"

--share-data explicitly enables sharing. Contribute only content you have permission to share. Redaction is best effort, records are retained until deleted, and billing is unchanged. Read the privacy policy. Background --document jobs do not participate.

One whole document

With a funded workspace key in CLASSIFY_API_KEY, upload a UTF-8 document of up to 10 million tokens (100 MB):

classify renewal,cancellation --document agreement.txt --json

The CLI streams the file once and waits for the result. The server handles chunking, screening, retries and final judgment. The price is $0.084 per million original tokens; 10M tokens costs $0.84. Failed jobs are refunded. CLASSIFY_JOB_TIMEOUT controls how long to poll (seconds, default 86400). Interrupting the CLI does not cancel accepted work; stderr includes its job ID.

Built for agents

The confidence is calibrated (on a six-way emotion set, answers at ≥ 0.9 were right 82% of the time; below 0.5, 29%), so the useful move is to trust the sure ones and look at the rest yourself:

classify relevant,"not relevant" -i "relevant means about GPU pricing" --review 0.7 < snippets.txt

prints only the inputs the model was unsure about. Or let the API do it: --smart re-asks those of a reasoning model.

Input can be plain lines, a JSON array, or NDJSON; --field picks the text field and --id carries an id through:

cat issues.jsonl | classify bug,question,feature --field title --id number --json

10,000 fast-tier lines are batched 1,000 per request, four at a time. Smart requests without a partner key use batches of at most 200 and one worker, so the 200/minute quota is not spent by concurrent requests. A minute rate limit pauses and resumes; a daily quota stops immediately. Errors go to stderr with exit code 1.

Options

-m, --multi                every label that applies, plus a score per label
-k, --max <n>              at most n labels (implies --multi)
-s, --smart                re-ask uncertain answers of a reasoning model
    --no-share-data       opt out of training-data retention
    --share-data          explicitly enable sharing (default)
-i, --instructions <text>  extra criteria
    --document <file>      one whole UTF-8 document (funded workspace)
-r, --review <t>           print only inputs with confidence below t
-c, --count                label histogram instead of rows
-j, --json                 NDJSON output
-q, --quiet                labels only
    --field <name>         text field for JSON / NDJSON input (default: text)
    --id <name>            id field to carry through
    --endpoint <url>       API base (env CLASSIFY_ENDPOINT)
    --api-key <key>        bearer token for higher limits (env CLASSIFY_API_KEY)

classify --help has examples. Environment: CLASSIFY_BATCH (a whole number from 1 to 1,000; public smart is capped at 200), CLASSIFY_TIMEOUT (seconds per request, default 180), CLASSIFY_NO_PROGRESS, and CLASSIFY_NO_UPDATE_CHECK=1 to skip the once-a-day version check.

What it refuses to do quietly

An answer from the API that is short, empty, or in the wrong shape stops the run with exit 1 rather than printing fewer rows. Every retry — rate limit, upstream error, network — is announced on stderr, and a timed-out request is retried once, not five times. With --smart, answers the reasoning model could not re-ask are counted on stderr, so a degraded smart tier never looks like a normal one.

Versioning

Semver. classify --version prints it; the CLI mentions a newer release on stderr once a day. Changes are in CHANGELOG.md; releases are cut with npm run release (see release.js) and published from the cli-v* tag.