Commands

Every openlabel command, what it needs, and how it exits.

Command	Does	Needs an API key
check	Judges a piece’s claims against its label	Yes, any provider
watch	Lists saved claims a label change touches	No
unseen	Finds prescribers with little industry contact	Only for a question, any provider
search	Prints the label sentences matching a question	No
diff	Prints what changed between two label versions	No
skills eval	Runs the skills’ eval suites	Yes, Anthropic

Every command takes --json to print machine-readable output instead of text, and -h or --help.

Color is on when output goes to a terminal. Set NO_COLOR to turn it off.

Models and providers

Check and the plain-language form of Unseen work with any of these. Open Label uses the provider you name with --provider (or OPENLABEL_PROVIDER), otherwise the first one whose key is set, in this order:

Provider	Key	Model
anthropic	ANTHROPIC_API_KEY	Optional. Defaults to claude-opus-5-5
openai	OPENAI_API_KEY	Required
openrouter	OPENROUTER_API_KEY	Required
gemini	GEMINI_API_KEY	Required
groq	GROQ_API_KEY	Required
ollama	none	Required; runs against http://localhost:11434

Set the model with --model <id> or OPENLABEL_MODEL. Every provider except Anthropic needs one, because Open Label can’t know which of a provider’s models you have access to. OPENLABEL_BASE_URL points any OpenAI-style provider at another address, such as a company gateway.

openlabel check email.html --label dupixent --provider openrouter --model <model id>

Open Label asks every model for JSON that matches a schema and rejects answers that don’t. The guards on verdicts apply whatever the model: a claim must quote the piece, and a citation must be a sentence the model was shown. How well a model judges claims still varies, so run a few pieces you know before relying on one.

skills eval always runs Claude, because the skills are written for Claude Code.

A .env file in the directory you run from is loaded automatically.

Exit codes

Code	Meaning
0	Success with nothing to report
1	Something needs attention: an overstated claim, an affected piece, a failed eval case
2	check only: an off-label or unsupported claim, or a rule violation
64	Usage error
69	Could not run: network or API failure

See Verdicts and exit codes for exactly when check returns each.