@creatorline/cli

Generate AI influencer photos and videos from your terminal, your CI, and your coding agent.

npm install -g @creatorline/cli
crl auth login          # paste a key from Settings → API
crl gen create text-to-photo --prompt "golden hour rooftop" --wait --output-dir ./out

Zero runtime dependencies, Node 20+. Full docs: docs.creatorline.io · Help: [email protected]

Grammar

crl <command> <subcommand> [args] [flags]

| Command | Subcommands | | --- | --- | | auth | login, logout, whoami | | models | list [tool], get <model-id> | | voices | list | | accounts | list, use <id\|name>, current, clear | | gen | create <tool>, cost <tool>, get <id>, wait <id>, list | | posts | create, publish <id>, list, get <id> | | assets | list, get <id> | | upload | crl upload <file> | | credits | — | | mcp | config, install |

generate aliases gen, creators aliases accounts, clips aliases posts.

Pick a creator once

crl accounts list
crl accounts use Eda        # persisted, like the studio's creator switcher

Every later gen create and posts create runs as that creator — the output lands on their wall and posts go to their connected channels. --account-id overrides it for one command.

Generate, then publish

ASSET=$(crl gen create text-to-photo --prompt "matcha latte flatlay" --wait --json | jq -r '.output[0].asset_id')
crl posts create --asset "$ASSET" --caption "morning ritual ☕️" --publish now

posts create alone leaves the post in the studio's review queue (pending_review) — nothing is public. --publish now|schedule|queue chains the publish call, and crl posts publish <id> --now does it later. Publishing to a live account asks for confirmation on a TTY; pass --yes in scripts.

crl posts list --status published
crl posts get <clip-id>          # per-channel status and the live URLs

Global flags

| Flag | Default | | | --- | --- | --- | | --json | | machine-readable output, and nothing else on stdout | | --wait | | block until the generation finishes | | --wait-timeout | 10m | give up waiting (the run continues server-side) | | --wait-interval | 3s | poll interval | | --output-dir | | download results to disk | | --api-key, --api-url | | override the stored credentials | | --no-color, --help, --version | | |

Every flag accepts both spellings: --aspect-ratio and --aspect_ratio are the same flag. Model parameters are snake_case and CLI convention is kebab-case; guessing wrong should not be a failure mode.

The model's flags are the model's schema

crl gen create fetches the model's schema first, then interprets the remaining flags against it. So the CLI never hardcodes a model's parameters, and a new model works the day it ships:

crl models get bytedance-seedance-2-0-reference-to-video-fast
crl gen create ugc-video --prompt "unboxing" --image ./face.png --duration 5 --aspect_ratio 9:16 --wait

models get prints the whole schema: each input's modality and cap, each parameter's allowed values or its numeric bounds (--duration 4–60 step 1), what opaque values like caption styles and scene modes mean, whether a prompt is required, and the model's maximum take length.

A wrong flag tells you the fix:

error Unknown flag --durations for model bytedance-seedance-2-0-reference-to-video-fast.
  inputs:     --image_urls --source_video
  parameters: --duration --aspect_ratio --resolution --generate_audio
  Run `crl models get bytedance-seedance-2-0-reference-to-video-fast` for values and defaults.

Run flags that are not model parameters

| Flag | | | --- | --- | | --source-url <url> | work from a video link instead of a file, for a tool that takes one (short-clip). A bare URL positional does the same. | | --no-account-avatar | keep the selected creator out of the output — a product shot posted on their account. | | --voice_id, --voice_a, --voice_b | a voice from crl voices list; without one, the creator's own cloned voice is used. |

crl gen create short-clip https://youtu.be/xyz --wait --output-dir ./shorts
crl voices list --scope library --search "warm female" --language en
crl gen create text-to-speech --prompt "hey, welcome back" --voice_id <voice-id> --wait

Media flags take a path OR an id

--image ./face.png uploads the file and uses it. --image <asset-id> reuses an existing asset. --image https://… hands the URL over as-is. No separate upload step, no juggling ids — though crl upload <file> exists when you want to pin one.

Prompts from stdin

echo "a cinematic city at dusk" | crl gen create text-to-photo --wait

Cost before you spend

crl gen cost ugc-video --duration 10        # → "48 credits (…, balance 1216)"
crl credits                                 # balance + recent ledger

lipsync, motion-control, scene-video, change-voice and short-clip are billed per second of the source, not of the output. Give gen cost the same media the real run will use — an asset id, a URL or --source-url — and it prices the actual length:

crl gen cost short-clip https://youtu.be/xyz   # → "32 credits (short-clip, 635s, …)"

Failed and moderation-blocked runs are refunded automatically.

Scripting and agents

Results go to stdout, progress and errors to stderr, so pipes stay clean:

crl gen list --json | jq -r '.data[] | select(.status=="succeeded") | .output[0].url'

With --wait, crl gen create prints the media URL(s). Without it, the generation id.

Parallel fan-out is just shell job control — --wait blocks per process:

for tool in text-to-photo ugc-video; do
  crl gen create "$tool" --prompt "$PROMPT" --wait --output-dir "./out/$tool" &
done
wait

Exit codes: 0 success · 1 usage or runtime error · 2 not authenticated · 3 out of credits.

Configuration

Precedence: flags → environment → config file.

| | | | --- | --- | | CREATORLINE_API_KEY | the key | | CREATORLINE_API_URL | the API root (default https://api.creatorline.io; a dev server is http://localhost:3000/api) | | CREATORLINE_APP_URL | the studio, for printed links (default https://app.creatorline.io) | | ~/.config/creatorline/config.json | written by crl auth login, mode 0600 |

CI wants the environment variable; there is no interactive login there.

MCP

crl mcp config     # print the snippets
crl mcp install    # run `claude mcp add` for you

The MCP server is remote and stateless — the CLI does not host it. Result URLs from stored objects are time-limited; download what you need.