@loopstack/cli

Scaffold, run, inspect, and watch Loopstack workflows from the terminal — live transition and LLM token streaming, human-in-the-loop prompts answered inline, and CI-ready JSON output with meaningful exit codes.

npm install -g @loopstack/cli

Start a new app with a zero-config hello workflow:

loopstack create my-app
cd my-app && docker compose up -d && npm run start:dev

$ loopstack run hello --arg name=Maya

▸ run 088cd19c… started
▸ start
✓ start (9ms)
Hello, Maya! 👋
result: {"greeting":"Hello, Maya! 👋"}
■ run completed in 96ms

When a run asks a question, answer it right there — or in Studio; whichever answers first wins, and the CLI keeps following either way. Choices, confirms, chat inputs, and secret entry render as native prompts; forms submit as-is or open the complete content in $EDITOR first:

⏸ waiting_for_answer

Which environment should we deploy to?
  1. staging
  2. production
answer: 2
■ run completed

Input only Studio can collect is named explicitly (with a deep link) instead of hanging, and browser-based waits — like an OAuth sign-in link — resume the attached session automatically once completed.

Commands

| Command | What it does | | --------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | loopstack create <dir> | Scaffold a new Loopstack app (NestJS + LoopstackModule + hello workflow + docker-compose) | | loopstack list [workflow] | Workflows you can run; with a name: its arguments and a ready-to-run example | | loopstack run <workflow> | Start a run and follow it live; --arg k=v, --arg [email protected], --arg k=@- (stdin), --detach, --quiet, --open, --json | | loopstack runs [run-id] | Recent runs (waiting-for-input first) or one run’s full transcript; --json includes the pending prompt (pendingPrompt), --record <file> derives a replay fixture from the run’s recorded tool calls | | loopstack answer <run-id> | Answer a waiting run’s prompt non-interactively — --arg k=v fields or --payload <json>; for agents and scripts | | loopstack attach <run-id> | Rejoin a run live — transcript so far, then streaming output and interactive prompts (docker/tmux semantics) | | loopstack watch | Live event stream; --workflow, --type, NDJSON with --json | | loopstack login | Save a backend environment to ~/.loopstack/config.json | | loopstack env list\|use | Manage saved environments |

CI

- run: npx --yes @loopstack/cli run triage-ticket --arg ticket=@samples/ticket-042.json --json
  env:
    LOOPSTACK_URL: ${{ vars.LOOPSTACK_URL }}
    LOOPSTACK_TOKEN: ${{ secrets.LOOPSTACK_TOKEN }}

Exit codes: 0 completed · 1 failed · 2 connection/config error · 3 waiting for input (non-interactive). --json prints one final object — { workflowId, status, result, errorMessage, durationMs } — on a clean stdout.

Local backends with auth disabled need no login or token at all. When an environment has a Studio URL (saved via login, LOOPSTACK_STUDIO_URL, or the local default), run output includes deep links into Studio — and --open jumps straight there. Full documentation: CLI Reference.