agency-os is an open-source plugin marketplace for Claude Code built for web & advertising design agencies. It packages 40 skills, 9 commands, and 17 specialist agents into 6 installable bundles that cover the entire agency lifecycle — from the first client call to the legal go-live gate and the monthly retainer loop. One workflow router keeps every project on rails: BRIEF → CONCEPT → DESIGN → BUILD → QA → LAUNCH → MAINTAIN.
Built by an agency, for agencies. Everything here runs in real client production — bundled from the best open-source skills (fully attributed, original licenses kept) plus original skills written for agency work.
- What is this?
- Quick start
- The project pipeline
- Phase gates
- The six bundles
- Commands
- Agents
- Walkthrough: a real project
- Bring your own designs
- FAQ
- Recommended companion plugins
- Credits
- Contributing
- License
If you build websites or run ad campaigns for clients with Claude Code, you have probably collected skills — a GSAP skill here, an SEO skill there, a design skill from a blog post. Two problems show up fast:
- Trigger chaos. With 31 separate SEO skills installed, "do an SEO audit" fires a random one. With 8 GSAP skills, nobody knows which loads.
- No process. Skills are ingredients, not a kitchen. Nothing connects the client brief to the design, the design to the build, the build to the launch.
agency-os fixes both:
- Consolidation with trigger discipline. Every skill description states what it does and what it does not do — and names the sibling skill that covers the neighboring case. "SEO audit" hits exactly one skill. Every time.
- A workflow spine. The
agency-workflowrouter walks every client project through six phases with hard exit gates, loading the right specialist skill at the right moment — automatically.
You install only the bundles you need. Each is a standard Claude Code plugin.
1. Add the marketplace (one time, inside Claude Code):
/plugin marketplace add gekkos-tech/agency-os
2. Install bundles — all six, or just what you need:
/plugin install agency-core@agency-os
/plugin install design-foundation@agency-os
/plugin install motion-3d@agency-os
/plugin install quality-qa@agency-os
/plugin install seo-marketing@agency-os
/plugin install ad-creative@agency-os
3. Start your first project (new session so the plugins load):
/new-project Café in Cologne, modern-minimalist, online reservations
That's it. Claude runs the client intake, writes docs/brief.md, and tells you
what the next phase needs. No configuration, no API keys, no accounts.
Minimal install: if you only want the workflow,
agency-corealone works — it degrades gracefully and tells you which specialist bundles would help.
agency-workflow (in agency-core) is a router: it detects which phase your
project is in, loads the matching specialist skills, and refuses to advance until
the phase's exit gate passes. Here is what each phase actually does:
| Phase | What happens | What you get | Specialists loaded |
|---|---|---|---|
| 1 · BRIEF | Client intake interview — goal, audience, scope, content, constraints, taste anchors | docs/brief.md the client could sign |
— (built into the router) |
| 2 · CONCEPT | Sitemap, a job + narrative for every page, creative direction, conversion path | docs/concept.md |
taste, awesome-design-md, ad-concepts |
| 3 · DESIGN | Commit to ONE style system, generate one reference image per section, assemble the binding handoff | design/ folder: tokens + section images + notes |
style-systems, ui-ux-pro-max, imagegen-web, brandkit |
| 4 · BUILD | Plan before code. Implement the handoff 1:1 — the handoff wins every argument | Working code, per-increment file lists | image-to-code, gsap-suite, interface-polish, karpathy-guidelines |
| 5 · QA | Evidence-based verification: function, fidelity, performance, accessibility | docs/qa-report.md |
performance, core-web-vitals, accessibility, webapp-testing |
| 6 · LAUNCH | Legal + technical go-live gate, SEO baseline, cutover plan, handover | docs/launch-checklist.md with GO/NO-GO |
pre-launch-checklist, seo-audit, social-formats |
| 7 · MAINTAIN | The post-launch loop: drift checks after deploys, monthly CWV/function re-tests, client reports | docs/reports/YYYY-MM.md per cadence |
seo-audit (drift), core-web-vitals, webapp-testing |
Every project also carries a docs/project-state.md — phase, gate status,
blockers with owners, who owes what, and the single next step. The router
updates it at every phase end; /status reads it back. That's how a project
survives session boundaries and teammate handoffs.
Two principles run through everything:
- The handoff is the truth. Once
design/exists, no skill regenerates colors, typography, or layout. Quality skills review; they don't redesign. - Plan before code. Every build increment states what will be built, from which design source, touching which files — before any code is written.
Each phase ends with a checklist that must pass. A few examples of what the
gates catch: a brief with three "primary" goals (force-ranked to one), a page
with no CTA, a handoff a developer couldn't build from, a form that submits but
never emails anyone, and a site about to launch with staging noindex still set.
The workflow router plus code-discipline guards. Install this first.
| Skill | What it does |
|---|---|
agency-workflow |
Routes BRIEF → CONCEPT → DESIGN → BUILD → QA → LAUNCH → MAINTAIN with per-phase references, exit gates, and the project-state file |
complete-output |
Bans placeholder patterns ("// rest stays the same"), enforces complete deliverables, clean split protocol |
karpathy-guidelines |
Behavioral guardrails against common LLM coding mistakes — surgical changes, surfaced assumptions |
cms-integration |
Which CMS (if any): decision tree + integration patterns for Payload, Sanity, Storyblok, Strapi, Decap, headless WordPress — schema modeled from the design handoff, editor-proof constraints, handover checklist |
Commands: /new-project · /handoff · /proposal · /status · /monthly-report
Everything that makes output look designed instead of generated.
| Skill | What it does |
|---|---|
ui-ux-pro-max |
Searchable design intelligence: 50+ styles, 161 palettes, 57 font pairings, 99 UX guidelines, 10 stacks |
taste |
Anti-generic creative direction from a brief — reads the room before choosing an aesthetic |
impeccable |
Command-driven design workflow: craft, audit, polish, animate, harden |
style-systems |
Commits a project to ONE of four complete style systems (minimalist-editorial, brutalist-industrial, soft-depth, product-modern) — token-level, never mixed |
redesign |
Audit-first premium upgrade of existing sites without breaking functionality |
interface-polish |
The last 10%: hover states, easing, layered shadows, optical alignment, tabular numbers |
awesome-design-md |
74 real brand design systems (Linear, Stripe, Apple…) as ground-truth DESIGN.md references |
Command: /design-review + design-reviewer agent
| Skill | What it does |
|---|---|
gsap-suite |
The 8 official GSAP skills in one: core tweens, timelines, ScrollTrigger, plugins, React, Vue/Svelte, utils, performance |
threejs-webgl |
Three.js scenes, WebGL/WebGPU, configurators, 3D visualization |
react-three-fiber |
Declarative 3D in React |
lottie-animations |
After-Effects vector animation on the web |
rive-interactive |
State-machine animation with runtime interactivity |
remotion-video |
Programmatic video rendering in React (not browser animation — actual video files) |
| Skill | What it does |
|---|---|
performance / core-web-vitals |
Load-time optimization; LCP, INP, CLS diagnosis and fixes |
accessibility |
WCAG 2.2 audits and fixes |
best-practices |
Security, compatibility, code quality |
webapp-testing |
Playwright-driven functional testing of local web apps |
pre-launch-checklist |
The German/EU go-live gate: DSGVO, Impressum (§5 DDG), cookie consent (TDDDG), BFSG accessibility, self-hosted fonts — plus SSL, redirects, 404, OG tags, sitemap, forms. Legal items are hard blockers |
Command: /pre-launch + qa-auditor agent
claude-seo v2.2.0, consolidated from 31 skills into 9 so triggers never collide — plus 15 specialist subagents for parallel audit delegation.
| Skill | What it does |
|---|---|
seo-audit |
The single audit entry point: crawls up to 500 pages, delegates to specialists, 0-100 health score, action plan, PDF report. Includes single-page audits, drift monitoring, Unlighthouse |
seo-technical |
Crawlability, robots, canonicals, redirects + sitemaps, hreflang, image SEO |
seo-content |
E-E-A-T, content briefs, SERP-overlap topic clustering, programmatic SEO, SXO |
seo-schema |
JSON-LD detection, validation, generation |
seo-local |
Google Business Profile, NAP, citations, map pack + geo-grid maps intelligence |
seo-ai-visibility |
GEO: AI Overviews, ChatGPT search, Perplexity citations, llms.txt, FLOW framework |
seo-competitor |
"X vs Y" and alternatives pages with comparison schema |
seo-backlinks |
Link profiles via free sources (Common Crawl) — no paid API needed |
seo-ecommerce |
Product schema at catalog scale, Google Shopping, marketplace intelligence |
No API keys required for any of it. If you have keys (DataForSEO, Ahrefs, Google APIs…), the upstream claude-seo marketplace adds those power-ups.
| Skill | What it does |
|---|---|
ad-concepts |
Campaign ideation: always exactly 3 distinct routes (claim, hook, visual idea, channel) built on AIDA/PAS/JTBD/4U + a hook formula bank |
ad-copy |
Performance copywriting in German AND English: headline systems, funnel-stage CTA library, A/B/C protocol with hypotheses, Du/Sie decision tree, UWG/HWG compliance notes |
social-formats |
Platform specs for Meta/IG, TikTok, LinkedIn, YouTube — verified July 2026 — plus a master-creative adaptation workflow with safe zones |
landingpage-cro |
8-section conversion narrative, message-match rule, friction audit, DACH trust elements |
brandkit |
Art direction for AI-generated brand boards, logo sheets, identity decks |
imagegen-web |
One reference image per page section — the design source developers rebuild from |
imagegen-mobile |
Premium app-screen concepts with multi-screen token consistency |
image-to-code |
Faithful implementation FROM design images: deep analysis pass, fidelity rules, verification against every source image |
Commands: /campaign · /adapt-formats
| Command | Bundle | What it does |
|---|---|---|
/new-project <brief> |
agency-core | Kick off a client project — runs the BRIEF phase |
/handoff [path|verify] |
agency-core | Assemble the binding design handoff from your assets, or audit it for completeness |
/proposal [terms] |
agency-core | Client proposal from the brief: scope, phases, deliverables, assumptions, out-of-scope list |
/status [path] |
agency-core | Where does the project stand? Phase, gates, blockers with owners, the single next step |
/monthly-report <url> |
agency-core | The MAINTAIN loop: drift check, CWV spot-check, function re-test, client report |
/design-review [target] |
design-foundation | Structured review: verdict, fidelity deviations, findings by severity, quick wins |
/pre-launch <url> |
quality-qa | The full legal + technical go-live gate with GO/NO-GO verdict |
/campaign <client + goal> |
ad-creative | 3 campaign routes with different emotional registers |
/adapt-formats <asset> |
ad-creative | Format matrix + adaptation plan for every placement in your media plan |
Two review agents ship with the bundles — design-reviewer (design-foundation)
and qa-auditor (quality-qa) — plus 15 SEO specialists (seo-marketing) that
seo-audit dispatches in parallel: technical, content, schema, performance,
visual, local, maps, GEO, backlinks, clustering, SXO, drift, e-commerce, FLOW,
sitemap.
What actually happens when you type:
/new-project Café in Cologne, modern-minimalist, online reservations
- BRIEF — Claude extracts what it already knows (business: café; market:
Cologne; scope hint: reservations), asks the missing intake questions in one
batch (goal ranking, content inventory, taste anchors), and writes
docs/brief.md— in German, because the client is German. - CONCEPT — sitemap (Home, Menü, Reservierung, Über uns, Kontakt), a
narrative per page, creative direction via
taste("warm minimalism, no generic café clichés"), one CTA per page (Reservieren). - DESIGN —
style-systemscommits to minimalist-editorial, tokens get written down,imagegen-webproduces one reference image per section, and/handoffassembles the bindingdesign/folder. - BUILD —
image-to-codeimplements section by section against the images;gsap-suiteadds the scroll entrance; scope stays strict. - QA — forms actually submit, Core Web Vitals measured, keyboard path
tested;
docs/qa-report.mdwritten. - LAUNCH —
/pre-launchruns the DSGVO/Impressum/consent/BFSG gate (reservation systems are BFSG-relevant!), SEO baseline viaseo-audit, cutover plan, handover doc. GO.
Already have designs (from Figma, a designer, or a previous Claude session)? They become the binding source of truth — nothing gets regenerated:
/new-project Café in Cologne, online reservations.
Designs are in ./designs/ — they are binding.
or, mid-project:
/handoff ./designs/
The DESIGN phase then skips generation entirely: tokens are extracted from
your files into design/tokens.md, your images become design/sections/, and
BUILD implements them 1:1 — with a verification pass against every source image.
Do I need all six bundles?
No. They're independent plugins. agency-core alone gives you the workflow;
it detects which specialist bundles are installed and degrades gracefully.
Does this cost anything? Do I need API keys? The marketplace is free (MIT) and nothing in it requires an API key or account. Skills that would need paid APIs were deliberately left out (linked under companions instead).
Why is there German legal stuff in my QA bundle? Because forgetting Impressum, cookie consent, or BFSG accessibility is how German agencies get Abmahnungen. If you don't serve DACH/EU clients, those checklist items simply won't apply — the technical half (SSL, redirects, sitemap, OG tags) is universal. Note: the checklist is a practical aid, not legal advice.
Does it work in German?
Yes — triggers respond to German phrasing ("Neues Projekt", "Kampagne für",
"SEO-Audit"), client-facing documents follow the client's language, and
ad-copy writes German and English copy natively (including the Du/Sie
decision).
How is this different from installing the upstream skills directly? Three things: consolidation (31 SEO skills → 9, 8 GSAP skills → 1 — so triggers don't collide), the workflow spine connecting them, and curation (every bundle skill works without keys, carries its license, and has a description that says when not to use it).
Can I use only the SEO part? Only the design part? Yes — that's the point of bundles. Each plugin stands alone.
Something triggered wrong or didn't trigger. What do I do? Open an issue with the exact prompt you typed. Trigger descriptions are the most-tuned part of this repo and we treat misfires as bugs.
Not bundled (different licenses or they need accounts/keys), but they pair well:
- frontend-design — Anthropic's official frontend design plugin
- anthropics/skills — official skills incl. document handling
- Vercel plugin — deploys, envs, previews from Claude Code
- claude-seo — the upstream SEO marketplace; adds API-powered extensions (DataForSEO, Ahrefs, SE Ranking, Profound, Bing, Firecrawl, Google APIs, image generation)
agency-os stands on excellent open-source work. Bundled skills keep their original licenses and LICENSE files — full details in THIRD_PARTY_LICENSES.md:
- ui-ux-pro-max — Next Level Builder (MIT)
- taste — Leonxlnx (MIT)
- impeccable — Paul Bakaus (Apache-2.0)
- awesome-design-md — VoltAgent (MIT)
- GSAP skills — GreenSock/Webflow (MIT)
- claude-seo — Daniel Agrici and Pro Hub Challenge contributors (MIT); FLOW framework prompts (CC BY 4.0)
- web-quality-skills — Addy Osmani (MIT)
- webapp-testing — Anthropic (Apache-2.0)
- claudedesignskills — freshtechbro (MIT)
- karpathy-guidelines — community skill (MIT), inspired by Andrej Karpathy's writing
Original skills (agency-workflow, style-systems, redesign, interface-polish, complete-output, remotion-video, brandkit, imagegen-web/mobile, image-to-code, ad-concepts, ad-copy, social-formats, landingpage-cro, pre-launch-checklist) © 2026 GEKKOS Tech, MIT.
Issues and PRs welcome. Ground rules:
- Licensing first — third-party content only with a clear MIT/Apache/CC license, original LICENSE bundled, entry in THIRD_PARTY_LICENSES.md.
- Trigger discipline — new skill descriptions must state when to use them AND which sibling skill covers the neighboring case.
- Skill format — SKILL.md < 500 lines, details in
references/, valid frontmatter (name= directory name). - CI must stay green — every push runs
claude plugin validate --strictplusscripts/validate_frontmatter.py(frontmatter, manifests, reference links). Behavioral eval cases live inevals/and run locally viaclaude plugin eval(early access).
MIT © 2026 GEKKOS Tech. Bundled third-party skills retain their original licenses — see THIRD_PARTY_LICENSES.md.