Three packaged Agent Skills for working with Krea.ai. Install once, get the Krea skill suite.
| Skill | When to use |
|---|---|
krea-generate |
Canonical generic-generation router. Image, video primitives, enhancement, edits, LoRA, portraits, text/poster work, and archviz through connected Krea MCP tools. |
krea-marketing |
Marketing creative workflow: product photoshoots, marketplace image sets, DTC static ad templates (one product photo → a library of on-brand ad formats), key visuals, UGC/social ads, campaign packs, and optional Meta Ads performance context. |
krea-motion |
Product and brand motion: cinematic shots, multi-beat reveal films, logo stings, luxury and beauty product loops, edit assembly, QA, and retakes. Not for anime or narrative animation. |
Experimental app-integration material lives under wip/ and is not installed or packaged by default.
Works with Claude Code, Cursor, Codex, Windsurf, OpenCode, Gemini, OpenClaw, and any agent that picks up ~/.<agent>/skills/<name>/SKILL.md.
npx skills add krea-ai/skillsInstalls the packaged skills. This is the supported install path across agents that use skill packages.
The packaged Codex and Cursor plugins use a connected Krea MCP server.
Check your agent's MCP tool list for Krea tools. If the tools are missing or auth has expired, connect or authenticate the Krea MCP server before using generation skills. Cursor plugin installs discover the remote server from the root mcp.json and start Cursor's OAuth flow; no API key or client secret is stored in this repository. For Codex plugin installs, the practical reauth path is to uninstall and reinstall the Krea plugin so the install auth flow runs again.
This is a single-plugin repository with provider-specific manifests that share the same three skills:
- Cursor reads
.cursor-plugin/plugin.json, whose explicitskillspaths point at the existing top-level skill directories, and discovers the Krea server frommcp.json. - Codex is packaged from
.codex-plugin/plugin.jsonand.codex-plugin/.mcp.jsonwithnpm run build:codex-plugin. - Claude reads
.claude-plugin/plugin.jsonand the shared.mcp.jsonconfiguration.
Validate the Cursor package before local testing or marketplace submission:
npm run validate:cursor-pluginFor local Cursor testing, symlink the repository to ~/.cursor/plugins/local/krea, reload Cursor, verify the Krea skills and MCP server in Customize, and complete the Krea OAuth prompt. The public repository can then be submitted at Cursor Marketplace Publish.
> Generate an image of a cyberpunk cat in neon rain.
> Make a 5-second video of waves at sunset.
> Upscale this photo to 4K.
These stay in krea-generate and route to the right model via the live MCP catalog.
> Render this Sketchup screenshot in golden hour, photoreal.
> Make my Revit model interior look photoreal, midday.
> Take this chair, put it in this room, in this style.
These route through krea-generate/workflows/archviz-3d-to-render.md for structural-reference archviz rendering.
> Make a hero product shot of my new perfume bottle.
> Generate a TikTok ad for my sneakers.
> Turn this one product photo into a set of on-brand static ad formats.
> Take this product URL and give me a marketplace full set plus social variants.
> Use Meta Ads performance context before making new ad creative.
These trigger krea-marketing. The agent starts with a compact creative intake. For paid-social, performance, campaign-analysis, catalog-performance, or activation work, it asks whether to connect performance context; otherwise it proceeds Krea-only from product refs, brand refs, and goals. Any live launch, budget, status, catalog, or publishing change stays gated and paused/draft by default unless explicitly approved.
The krea-generate skill includes an opt-in update notification. Once per session, the agent may run krea-generate/scripts/update-check.sh, which checks https://raw.githubusercontent.com/krea-ai/skills/main/VERSION against the installed krea-generate/VERSION file. If a newer version exists, it prints UPGRADE_AVAILABLE <local> <remote> to stdout - the agent surfaces this once, then continues. Snoozes 24h -> 48h -> 7d to avoid nagging.
To upgrade, re-run your install command (npx skills add krea-ai/skills, etc.).
To disable the check: touch ~/.krea-skills/update-check-disabled.
To force a fresh check: bash krea-generate/scripts/update-check.sh --force.
Single-source version: VERSION file at repo root. CI enforces that krea-generate/VERSION, all SKILL.md frontmatters, plugin manifests, and package.json agree.
Two layers live under evals/.
The headline Codex-plugin workflows, driven through real Codex (codex exec) against the live Krea MCP and graded outcome-based (an expensive op must not fire before the user approves) with an LLM judge for nuance. Details: evals/hero/README.md.
An offline gate runs on every push / PR — no secrets, no spend:
python evals/hero/run.py --lint # validate case specs + fixtures
python evals/hero/run.py --selftest # grader unit checksThe live suite spends real Codex + Krea + judge credits (~1–1.5h), so it is gated. It runs only on:
- a manual Run workflow dispatch of the Hero Evals action,
- a pull request labeled
run-hero-live, - a push to
maintouchingkrea-*/or.codex-plugin/.mcp.json, - a
repository_dispatch: mcp-changedfromkrea-ai/app.
⚠️ Caveat: a normal PR push (or a merge into the branch) does not run the live suite — you must add the exactrun-hero-livelabel, or dispatch it manually:gh workflow run evals.yml --ref <branch> # run the full real suite gh pr edit <pr-number> --add-label run-hero-live # or trigger via the PR label
46 local scenarios run through claude -p to catch routing / UX drift when skill content changes. Regex-graded (PASS / FAIL / MANUAL_REVIEW); see evals/README.md for methodology and the full scenario list.
bash evals/run.shMIT