SuperCMO Skills

Why it's ranked here

The engineering under the marketing copy is careful. Apache 2.0 throughout, a Python server with two third-party dependencies, and one routing layer that picks whichever vendor key you hold before falling back to the paid proxy. Long video jobs hand back a resumable job handle instead of timing out, and local reference files are refused unless their bytes are real image, video or audio, which stops a hijacked agent from shipping a credentials file to a vendor. The catch: the value lives in hosted models billed per generation, and usage telemetry is on until you switch it off.

What's good

The details show thought about how agents fail. Image and video calls accept a dry-run flag that returns the exact request with keys masked, so nothing costs money until you approve it. Up to ten images or clips go out in one call, and fan-out respects each vendor's concurrency ceiling: three parallel jobs on WaveSpeed's default tier, eight on fal. Telemetry sends a fixed list of fields, reduces raw errors to a normalized code, and its documentation names the endpoint and retention window. The ad copy skill refuses to invent ratings or testimonials and asks for real proof instead.

Tradeoffs

Nothing renders locally. Every image, clip and voice line is a paid call to a third-party model, so quality and cost track those vendors, and the promised shot-to-shot consistency rests on reference images passed to models the project does not control. Usage telemetry defaults to on for the server and posts to SuperCMO's own endpoint, which forwards to PostHog. Video lengths outside a model's allowed set are snapped to the nearest valid value rather than refused. Much of the craft is written instructions for the host agent, so results depend on how well that agent follows them. The package declares itself beta.

How to use it well

Suits a founder or small marketing team already working in Claude Code, Cursor or Codex who wants short product ads without juggling separate script, image, voice and video apps. Start with your own fal key, preview requests with a dry run, and approve the concept and script before any clip renders, since that is where money goes. Set one of the documented opt-out variables if usage events should not leave the machine. Posting to social accounts is a feature of the hosted product, not of this repository.

Observed

License

Apache-2.0 (LICENSE, declared in pyproject.toml and package.json)

Language

Python server and media client; Node.js installer

Runtime requirements

Python 3.11 or newer; Node 18 or newer for the installer

Packaging

PyPI-style wheel supercmo-skills and npm package @supercmo/skills

Interfaces

MCP server over stdio plus agent skills for Claude Code, Cursor and Codex

Third-party Python dependencies

icalendar and python-dateutil only

Generation backends

fal, WaveSpeed, ElevenLabs, Gemini, Firecrawl, ScrapeCreators, or SuperCMO managed proxy

Telemetry

Opt-out anonymous usage events with a field allowlist (telemetry.py, TELEMETRY.md)

Declared maturity

Development Status 4 - Beta in pyproject.toml

Cost preview

Dry-run mode on image and video generation returns the request without an API call