SuperCMO Skills

Creative production skills for AI agents

A collection of AI agent skills to research, plan, create, and ship full creative campaigns.

Describe what you want to make - like a UGC video, a product launch campaign, or a week of content. It plans the work, picks the best AI image/video/audio models, edits and delivers the finished, campaign-ready assets.

The open-source alternative to closed AI marketing agents.

Why SuperCMO Skills?

Quick start

Paste this to your coding agent (Claude Code, Cursor, Codex, …):

Run `npx --yes github:SupercmoHQ/superCMO-skills --all`. Tell me whether it succeeded, then follow the next steps it prints to set up a key.

Prefer to run it yourself:

npx --yes github:SupercmoHQ/superCMO-skills --all   # every detected host

Prefer a preview before spending? Every generation supports dry_run — the exact request and cost, no API call.

How the skills work together

You act as the creative director, and your agent uses SuperCMO skills to orchestrate the entire production pipeline.

Here is an example of how the skills chain together behind the scenes to execute a complex brief like: "Make a 1-minute video ad for this product [URL] with a voiceover."

[Brief] "Make a 1-minute video ad for this product [URL] with a voiceover."
   │
   ├── 1. Analyze ──> Scrapes the URL to understand the product and fetch product images.
   │
   ├── 2. Plan    ──> Breaks the 60s brief into four 15-second shots.
   │                  Writes the script and shot list.
   │
   ├── 3. Image  ──> Generates master reference images of the product 
   │                 using the best image models.
   │
   ├── 4. Video  ──> Generates the 4 video clips. Locks the anchor reference 
   │                  into every shot so the product doesn't shape-shift.
   │
   ├── 5. Audio   ──> Adds a voiceover, budgeting the script to match the 
   │                  exact length of the video.
   │
   └── 6. Stitch  ──> Stitches the finished clips together to create one 
                      continuous 1-minute video file.

Skills

These are the discrete capabilities installed into your agent.

Install

Every host loads the same skills + the same local MCP server - only the wiring differs. Prerequisite: uv on PATH — the MCP server runs via uvx, which provisions Python and the server for you (nothing to pip install). The root pyproject.toml / package.json are packaging scaffolding — you don't build anything by hand to run or contribute to the skills.

npx installer - registers the MCP server via each host's own mechanism (codex/claude mcp add where a CLI exists; a .cursor/mcp.json for Cursor) and places the skills. One command does every detected host:

npx --yes github:SupercmoHQ/superCMO-skills --claude               # Claude Code
npx --yes github:SupercmoHQ/superCMO-skills --cursor --project-dir .  # Cursor (per project)
npx --yes github:SupercmoHQ/superCMO-skills --codex                # Codex
npx --yes github:SupercmoHQ/superCMO-skills --all                  # every detected host

Claude Code plugin - an alternative to npx … --claude (use one, not both): the whole repo installs as one plugin, managed by Claude Code:

/plugin marketplace add SupercmoHQ/superCMO-skills
/plugin install supercmo@superCMO-skills

Codex plugin - an alternative to npx … --codex (use one, not both): installs the skills + MCP server as one plugin, managed by Codex:

codex plugin marketplace add SupercmoHQ/superCMO-skills
codex plugin add supercmo@superCMO-skills

Claude Cowork / Claude desktop - download supercmo-plugin.zip from the latest release, then Settings → Plugins → Upload local plugin.

Found us on skills.sh? npx skills add SupercmoHQ/superCMO-skills copies the skills but not the MCP server they call — so image_generate/video_generate won't exist yet. Install the server, then reload your agent:

npx --yes github:SupercmoHQ/superCMO-skills --all   # or --cursor --project-dir . · --codex · --all

Set up a key

Generation needs a key. Two ways — pick one:

Option A · Managed — one command

Generate on SuperCMO's keys, pay per use — no vendor signups, nothing to paste.

npx --yes github:SupercmoHQ/superCMO-skills login

Opens SuperCMO in your browser to sign in and authorize this device; the key is written to ~/.supercmo/.env automatically. Buy credits in the web app when you're ready.

Option B · Bring your own keys — free

Bring your own vendor keys — requests go directly to the model vendor; nothing routes through SuperCMO.

One file, every host. The installer creates ~/.supercmo/.env; the MCP server loads your keys from there on any host (Claude Code, Cursor, Codex, …). Open it, add a key, restart your host:

# ~/.supercmo/.env
WAVESPEED_API_KEY=your-key # image + video (start here) — https://wavespeed.ai
FAL_KEY=                   # image + video (alternative) — https://fal.ai
ELEVENLABS_API_KEY=        # voiceover (optional)        — https://elevenlabs.io
GEMINI_API_KEY=            # image/video analysis (opt.) — https://aistudio.google.com
FIRECRAWL_API_KEY=         # url extraction (optional)   — https://firecrawl.dev

• One key to start: WAVESPEED_API_KEY covers image + video and is what we recommend — users report better reliability there. FAL_KEY is fully supported too and serves the same models, so use it if that's the account you already have. With both set, WaveSpeed is used; SUPERCMO_MEDIA_PROVIDER=fal picks the other. The rest are optional — add one when you want that capability (voiceover, analysis, extraction).
• Prefer environment variables? Exporting the key (or putting it in your host's MCP config env block) also works and takes precedence — ~/.supercmo/.env is just the reliable default that also works for GUI-launched hosts that don't inherit your shell.
• Check what's set: ask your agent to run the setup_status tool (host-agnostic, no path needed).

Security & trust

These skills run inside your agent. Exactly what happens:

• Open source & inspectable. Every skill, script, and the MCP server lives in this repo under Apache-2.0 — read, diff, or pin before you run it.
• Your keys stay local (BYOK). With your own vendor keys, they live in ~/.supercmo/.env (or your host's MCP config env / your shell); the server reads them from its process environment and requests go directly to the vendor — nothing routes through SuperCMO. With the managed key, requests go to SuperCMO's proxy (billed to your credits) — you never hand us a vendor key.
• The MCP server is local + minimal. A stdlib-only Python package (supercmo-skills on PyPI, source in scripts/supercmo_skills/mcp/), fetched and run on demand via uvx supercmo-skills@<version> — it runs on your machine, launched by your host, and starts only when your host enables the plugin.
• Dry-run everything. Generation tools support dry_run - a free preview of the exact request (keys masked), no API call.

Found something off? Open an issue.

Telemetry

SuperCMO sends anonymous, opt-out usage counts from the MCP server so we can see which tools get used and prioritize. Full details in TELEMETRY.md.

• What we send: the tool name, whether it succeeded, how long it took, versions (OS / Python / SuperCMO), and a random install id. Nothing else.
• What we NEVER send: your prompts, tool arguments, generated media, file paths, keys, hostname, username, or IP address.
• Turn it off (any one): SUPERCMO_TELEMETRY=false, DO_NOT_TRACK=1, or DISABLE_TELEMETRY=1. It also honors Claude Code's CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC.
• See what would be sent: run with SUPERCMO_TELEMETRY=log - prints each payload instead of sending.
• Events go to our own endpoint (api.getsupercmo.ai), never a third-party analytics host. The random install id is never linked to any account.

Community

Questions, ideas, or something you built? Open an issue or discussion.

Contributing

New skills and improvements welcome - see CONTRIBUTING.md. In short:

1. Create skills/<your-skill>/ (folder name = the name in frontmatter), using skills/generating-images/ as the reference layout.
2. Keep SKILL.md short; push detail into references/, deterministic work into scripts/ (stdlib-only where possible, BYO-keys from env, --dry-run on anything that mutates).
3. Validate - CI runs the same on every PR:

python3 scripts/quick_validate.py       # structural + strict-YAML frontmatter (blocking)
python3 scripts/listing_gate.py         # scripts compile + --dry-run gates (blocking)
python3 scripts/check_shared_client.py  # no raw vendor HTTP - the brokering seam (blocking)
python3 scripts/check_catalog_sync.py   # provider-key catalog single-sourced (blocking)

License

Apache-2.0 - see LICENSE and NOTICE.

The skill files, scripts, and MCP server in this repository are Apache-2.0. The hosted SuperCMO product (getsupercmo.ai) is a separate service governed by its own terms.