CLI

npx startup-builder runs every operation from a terminal, a script, or a coding agent, with the same Noun.verb names and JSON input as the API. It also drains work: any Worker, human or agent, takes Tasks from it.

Sign inCopy Anchor Link

npx startup-builder login
npx startup-builder whoami

login opens the browser and stores a key. In CI, set STARTUP_BUILDER_API_KEY instead: the CLI uses the key in the environment first, then the stored one. whoami prints the identity, namespace, and scopes in use, which answers most access errors before they happen.

ReadCopy Anchor Link

The examples follow ledgerline.co, which closes the books each month for small online stores at $149 a month. Agents categorize every Transaction; a human bookkeeper reviews what they cannot.

npx startup-builder show /
npx startup-builder show /orders --fields '$id,total'
npx startup-builder search "stores with an open close"

show prints anything at an api.sb path: the Startup at /, a list, a record, or a definition. search finds records, Nouns, Verbs, and docs by keyword or meaning.

Run any operationCopy Anchor Link

npx startup-builder schema Order.refund
npx startup-builder do Order.refund --json '{"order":"order_31","amount":{"amount":149,"currency":"USD"},"reason":"Close missed the 5th"}' --try
npx startup-builder do Order.refund --json '{"order":"order_31","amount":{"amount":149,"currency":"USD"},"reason":"Close missed the 5th"}'

do takes the operation name and its input as one JSON object, the exact body the API takes. schema prints that input and the result's shape, so neither a person nor an agent guesses.

--try, --send, and --do pick the Mode, and writes default to --do. The first command above previews the refund and moves no money. The test environment is the default; --env live acts on live, and only with a key that reaches it.

Scripts and agentsCopy Anchor Link

The CLI switches to agent mode when its output is not a terminal, or when CI, CLAUDECODE, or CODEX is set. In agent mode it prints JSON, never prompts, and never opens a browser. A script reads the result from standard output and branches on the exit code:

A refused Action prints its problem document, with the fix and the next call. A coding agent reads both and retries without a search.

The work loopCopy Anchor Link

Work that needs a Worker waits as a Task. Four commands take it, finish it, or hand it back:

npx startup-builder next
npx startup-builder claim task_88
npx startup-builder done task_88 --json '{"result":{"categorized":14,"flagged":0}}'
npx startup-builder release task_88 --reason "needs the store's March bank statement"

next prints the next open Task the signed-in Worker's Role may take: here, 14 transactions the agent could not categorize. claim takes it with a 15-minute lease, so two Workers never do the same work. A lapsed lease returns the Task to the queue.

done finishes the Task with a result and resumes the Action that waited on it, so the close continues. release hands it back with a reason the next Worker reads first. pending lists the Actions the caller started that are still running.

Each command is one Verb, so the loop runs the same on all four Interfaces:

The bookkeeper and the categorizing agent drain one queue with the same four commands. Once the agent's answers match the bookkeeper's on 95% of a sample, that kind of Task stops reaching a person. The eval ladder sets the sample and demotes the agent on any regression.

create-startups is the command that creates a Startup.