skill

skills

High-standard open-source agent skills from OFOX AI — one monorepo. Install: npx skills add ofoxai/skills

ofoxai/skillscommunitylistedunreviewed

Set up

Skills are installed by placing their folder where your agent loads skills.

Download ZIP from GitHub Open the repository

The download comes from github.com/ofoxai/skills itself, not from our server, so what you get is exactly what its author published.

Install instructions come from the README

Skills do not use an MCP config. Follow the steps in the README below.

README

synced Oct 26, 2026

ofoxai/skills

High-standard, open-source agent skills from OFOX AI — one monorepo so every skill ships with the same quality bar (clear safety contracts, real-tool recipes, no leaking of secrets or local paths).

Skills work with Claude Code, Cursor, Copilot, and 70+ other agents via skills.sh.

简体中文

npx ofox-skills          # install every skill into every agent on this machine
npx ofox-skills doctor    # check which agents can actually see them

No account needed to price a job — see below. Read that part first: most of these skills spend real money.

The video and image skills cost real money — here's how to check before you commit

The video scenario skills — the four seedance-* ones plus keyframe-animation, product-demo, video-extend-edit, previs-rerender, ugc-ads, shorts-reels, talking-head, explainer and music-video — call Ofox's video API, which runs Seedance 2.5 and bills per second of generated video. A 15-second 720p clip runs about $3.60; a 4-second 480p draft is about $0.44, and there are cheaper models. Generation is a slot machine — you often want several takes and keep one — so the per-clip figure is not the whole cost.

The two image skills — image-edit and product-image — call the image API instead, which bills per output token, so there is no such thing as one price per picture: the same model has been measured 26x apart across two flags. Cents rather than dollars, but quote it rather than assume it. Two things make that worth doing properly. An edit also bills the picture you upload, so a large source costs more than a small one; and a set of four is four bills, which is why product-image quotes the set's total and never the per-image figure.

You can price any of this with no account and no API key. Install, then:

bash ~/.agents/skills/ofox-video-core/references/ofox-video.sh \
  generate --dry-run --prompt "two people arguing in a kitchen" \
  --duration 15 --resolution 720p
# Estimated cost: ~$3.60 (15s x $0.24/s)
# DRY RUN — nothing was submitted and nothing was billed.

~/.agents/skills/ is where the installer keeps the canonical copy, and every agent reads it — directly, or through a symlink of its own. Claude Code also exposes the same skill at ~/.claude/skills/ofox-video-core/, but that path only exists if Claude Code is installed, so the line above is the one that works everywhere.

The image side has the same escape hatch, including for an edit of a file you already have:

bash ~/.agents/skills/ofox-image-core/references/ofox-image.sh \
  edit --dry-run --image ./photo.jpg --prompt "replace the background with a beach"
# DRY RUN — nothing was submitted and nothing was billed.

--dry-run validates everything and quotes the price without sending a request. ofox-video.sh models / providers and ofox-image.sh models likewise need no key. Decide whether it's worth it, then sign up.

When you are ready: get a key at app.ofox.ai (Settings → API Keys → Create New Key, shown once), then

export OFOX_API_KEY=your_key_here

Already running Codex / Claude Code / Cline with an Ofox key configured? The same OFOX_API_KEY works — no new key needed.

Prerequisites: curl and jq. curl is usually preinstalled; jq often isn't (brew install jq on macOS, apt-get install jq on Debian/Ubuntu). The skills check for both and tell you what's missing.

hal-vault, hal-image and cloudflare-drop don't touch the Ofox API and cost nothing to run.

Every clip is recorded, and can be aimed at again

Generation is a slot machine, so the take you keep is worth being able to describe afterwards. Each finished clip lands as <name>-<short job id>.mp4 — pass --name and it is named after the scene instead of a bare job id — with a .json sidecar beside it holding the full job id, the prompt, the seed, the request as submitted, and what it actually cost.

The seed is the part that matters. Without one the server picks its own and reports it nowhere, so "that take was good, give me it at 1080p" has nothing to point at. Recorded, the whole request can be re-submitted at a different resolution.

What that is not is a reproduction. Measured 2026-09-15: three submissions of one byte-identical request on a fixed seed came back as two visibly different clips and one refusal. A re-render aims at a take; it does not return it — and these skills are written to say so before anyone pays for one.

To see what this looks like at the other end, the Seedance 2.5 prompts and examples page on ofox.ai publishes finished clips with the prompt and parameters behind each one — and, for the clips generated with these skills, the job id and the real bill.

Install

npx ofox-skills

Every skill, into every agent on your machine, at user level. That combination is the default because the other combinations break quietly:

Every agent, because skills add on its own asks you interactively which agents to install to — and when an agent is the one running it, skips the question and installs only to the agent it detects. Either way the agents you didn't pick get nothing, and you don't find out until one of them can't see a skill you know you installed.
Every skill, because each scenario skill reaches its execution layer by relative path, which only resolves when they sit side by side — the video scenarios need ofox-video-core, image-edit and product-image need ofox-image-core, and seedance-anime-drama needs both. The skills.sh manifest format has no dependency field to declare that with, so installing one alone can leave you with:
```
bash: ../ofox-video-core/references/ofox-video.sh: No such file or directory
```
That means the core skill is missing, not that the skill is broken.

Check what your agents can actually see

npx ofox-skills doctor

Lists every skill in this repo with the agents it is currently linked into, and exits non-zero if any are missing. Worth running when an agent insists a skill doesn't exist — usually it is right, and this says which ones and why.

It checks user-level (global) skills, which is where the default install puts them, and names that scope in its own output. If you installed with --project, ask for the same scope:

npx ofox-skills doctor --project

If everything is listed but an agent still can't see it, restart the agent; most read their skills once at startup.

Narrower installs

Any flag you pass overrides the matching default:

npx ofox-skills --agent codex               # one agent, still every skill
npx ofox-skills seedance-short-drama        # one skill, still every agent
npx ofox-skills --project                   # this project instead of user level

The underlying CLI works directly too, if you'd rather not go through this package — but then the defaults are yours to supply:

npx skills add ofoxai/skills --skill '*' --agent '*' --global --yes

Skills

Every skill in this repo, in three groups. The one-liners below are deliberately short — each SKILL.md carries the full contract, the flags, and the measured costs.

Video — Ofox video API (Seedance 2.5), bills per second

Skill	What it does
seedance-short-drama	Dialogue-driven scenes with real humans. Writes the shot list, quoted lines and delivery notes; one held take or several hard cuts inside one job.
seedance-ad-creative	Cinematic brand/product ads — hook, showcase, slow-motion climax, hero close. From a product photo or a text description.
seedance-product-video	Plain catalog/listing footage — white background, simple orbit or turntable, literal accuracy, no mood lighting. A product photo gives the best fidelity; a text description works for a generic or fictional product.
seedance-anime-drama	Anime/manga storyboard shots. Generates the character image first, then animates it, so the same character survives across shots.
keyframe-animation	Two stills you already have: image A as the first frame, image B as the last, in one job — the model fills the middle. Both ends come back honoured to the pixel — measured three times. When the subject gets there is not predictable: two runs produced two different easing curves.
product-demo	Two screenshots of your interface, before and after a state change, animated as one clip. The model cross-fades only the values that differ; the strings stay legible, so this one carries no anti-text rule.
video-extend-edit	Footage you already have, made longer — or given a different ending. A frame comes out of your clip locally for nothing and becomes the first frame of a new segment, which is joined onto the original. The new segment's pixel size comes from the tier you paid for and never from your clip, so the join has to rescale: concatenating without it produces a variable-resolution file that ffmpeg reports no error about.
previs-rerender	A rough white-model, clay, wireframe, blockout or animatic clip used as the blueprint for a finished treatment. One measured 4s/480p synthetic run kept its shot order, cut timing, framing, layout and motion directions; a second continued from the final state without replaying the rough cut. Other inputs and tiers remain experiments.
ugc-ads	Handheld, phone-shot creator clips — unboxing, first impression, honest review. Deliberately inverts the polish the other scenarios default to: one practical light, imperfect framing, no grading, no beauty filter.
shorts-reels	Several cheap vertical 9:16 drafts in one priced batch, a contact sheet to pick from, then one proper re-render of the winner. Brings the format and the economics; the prompt comes from whichever scenario skill fits.
talking-head	A portrait plus a short script, as one person saying those words to camera. You supply the text and the model generates the voice — audio can't be uploaded. The only skill here that doesn't default to Seedance 2.5: it refuses a real person's photo, so this one runs on `wan-3.0-prime`.
explainer	An article, doc or release note as a short spoken clip. A 30-second clip holds about eighty words in eight sentences — measured, and well under a tenth of a 1,200-word post — so it doesn't summarise the piece. Sentence count costs as much as word count, because each boundary is about 0.7s of silence. It picks the one idea worth saying, and helps you choose which.
music-video	Picture for a track you already have. The model never hears your audio — measured — so the visuals are written to the song's own sections, tempo and energy, generated as chained segments, and your file is laid on locally at the end for free. One job caps at 30 seconds, so a three-minute song is six jobs, and the cost table says so before anything is spent.
ofox-video-core	Library. The execution layer every video skill above calls: submit, poll, download, report the real cost. Install it, don't invoke it — unless you're driving the API directly.

Image — Ofox image API, bills per output token

Skill	What it does
image-edit	One change to a picture you already have — swap the background, recolour a part, remove an object — written as two sentences so what must not change is stated and checkable. One image in, one out.
product-image	A set of product images to choose between: several styles or backgrounds of one product, each an edit of the same photo so the item stays itself, priced as the set's total rather than per image.
ofox-image-core	Library. Every image model Ofox serves, `--model` defaulting to a cheapest-first chain that falls back and says so. Generates from text, or edits an image you already have (change the background, recolour an element) from a local file. Prices a job with `--dry-run`, then reports real token usage and dollar cost.

Free to run — no Ofox API, no per-call cost

Skill	What it does
hal-vault	SSH-key encrypted secret store. Masked by default, so an agent can search and inject a credential without ever printing it.
hal-image	ImageMagick recipes — resize, crop, composite, montage, watermark, convert — plus lossless compression before sending, so transfers stay small. Needs `magick` installed.
cloudflare-drop	A folder of static files to a live shareable URL. Permanent when `CLOUDFLARE_API_TOKEN` is set, a 60-minute claimable preview when not — and it says which one you got, verifies the served content, and refuses to invent a link. Optional `-otp` protects permanent links with a server-side six-digit access code.

Why a monorepo

One repo, one quality bar. Each skill is self-contained under skills/<name>/ (a SKILL.md plus optional references/), declared in skills.sh.json. Publishing many skills from a single high-standard repo is easier to govern, version, and review than a repo per skill — and consumers can still install any skill individually with ofoxai/skills@<name>.

See CONTRIBUTING.md for the bar every skill must clear.

ofoxai/awesome-seedance-2.5 — the Seedance 2.5 prompt and clip collection: official and community cases plus the ones generated with these skills, every case with its prompt and parameters, the self-run ones with the bill. The prompts page on ofox.ai is its on-site mirror.
ofoxai/hal-vault — the SSH-key encrypted secret store the hal-vault skill drives (Go CLI, built on age).