Stand up the brain, connect the tools, onboard the team, and start a client project. The team gets the work skills; the confidential layer stays with the founders. Secrets live in the host or keychain, never in Git.
crue-os/crue-os via the Claude GitHub App.crue-os/crue-house, mounted in the crue · House Project (GitHub connector, one way, read-only). The team repo carries no pricing. Edit in crue-house, push, then Sync.Everything is GitHub-backed. Each repo has one role, and one sync direction, so nothing drifts.
The engine: all skills, the shared standards, the Client Core template, the org marketplace. Public to the team, no pricing. Push here and the whole org gets it on sync.
The confidential layer: pricing, client data, proof. A separate private repo, mounted read-only in the crue House Project. The team gets every skill without it.
The Client Cores: one folder per client (clients/<slug>/) holding brand-pack, lock-ins, decisions. The Team project mounts it via Files → GitHub, selecting only that client's folder (read-only). Updated by commit (say "lock to client"), then Re-sync.
The actual deliverables: one repo per website or app (<client>-website), deployed on Vercel. The GitHub write MCP commits Cores and project files; it never writes crue-os or crue-house.
The org marketplace covers claude.ai chat and Cowork. Claude Code needs its own config, so teammates never run /plugin commands by hand. Set it once and the packs install and auto-update.
{
"extraKnownMarketplaces": {
"crue-os": { "source": { "source": "github", "repo": "crue-os/crue-os" }, "autoUpdate": true }
},
"enabledPlugins": {
"crue-brand@crue-os": true,
"crue-web@crue-os": true,
"crue-product@crue-os": true,
"crue-media@crue-os": true,
"crue-content@crue-os": true,
"crue-studio@crue-os": true,
"crue-ops@crue-os": true
}
}
Three ways to deploy it. User level (~/.claude/settings.json, paste once). Project level (.claude/settings.json committed into a crue-projects repo, so anyone who opens it gets the packs). Managed settings (MDM, zero-touch, org-wide).
Auth. Each user needs GitHub read access to crue-os/crue-os, the skills repo with no pricing. For background auto-update, set GITHUB_TOKEN (scope repo). autoUpdate keeps the packs current on startup.
Settings → Connectors → add, then sign in. These work in claude.ai web, Desktop and Code.
Figma · Canva · Sanity · Supabase · Vercel · ClickUp · Microsoft 365 · Higgsfield. Settings → Connectors → find each → Connect → authorize the account. Enable per Project where relevant.
The hands behind lock to client: it commits Client Cores to the clients repo. Add custom connector with URL https://api.githubcopilot.com/mcp/, Connect, sign in with GitHub.
When it asks which repositories, choose only crue-os/crue-clients (and crue-projects for site code), with Contents: read and write. Never grant write to crue-os or crue-house. Enable it in each client Project. Claude Code commits via native git, so the MCP is optional there.
Add custom connector with URL https://api.mobbin.com/mcp, then Connect and sign in with your Mobbin Pro account. Per-seat login.
Add custom connector with URL https://mcp.ideogram.ai/mcp, then Connect and sign in with your Ideogram account.
The real-vector engine: vector / SVG, icon sets, brand style sets, vectorize a raster to clean SVG, backgrounds and upscale. Added org-wide as a custom connector, Connect and sign in with your Recraft account (per-account login).
Add custom connector with URL https://bexio-mcp.crue.ch/mcp, Connect, then sign in with your crue Microsoft 365 account (Entra SSO). Enabled only in the crue · House Project.
Local stdio servers. Claude Code and Desktop only, by design. Per-user.
Component generation for the build. Get an API key from the 21st.dev Magic console, then:
npx @21st-dev/cli@latest install claude --api-key YOUR_KEY
Restart Claude, then use /ui in a build session. The key lives in local config, never in Git.
Captures reference sites (screenshots, structure, computed styles) and checks our own builds, the engine behind clone-from-reference and the web self-check. Ships bundled in the crue-web pack, so it auto-registers, no command to run. Needs Node 18+; the first run downloads a browser. Local only, so Code and Desktop, not the web, where the skills use the Chrome connector instead.
Real traffic and funnels, read-only. In Google Cloud: enable the Analytics Admin and Data APIs, create a Desktop OAuth client, then authorize and register:
gcloud auth application-default login \
--scopes=https://www.googleapis.com/auth/analytics.readonly,\
https://www.googleapis.com/auth/cloud-platform \
--client-id-file=YOUR_CLIENT.json
claude mcp add analytics-mcp --scope user \
-e GOOGLE_APPLICATION_CREDENTIALS=PATH \
-e GOOGLE_PROJECT_ID=PROJECT_ID \
-- pipx run analytics-mcp
Each teammate sets up their own Google login. It sees the GA properties that account can access.
The team gets the work skills. Founders also get the confidential layer and Bexio. Both stay current automatically.
Get added to the Claude org and the GitHub crue-os repo. Skills arrive in chat and Cowork automatically; for Claude Code, the settings.json above installs the packs. Open the right Project, describe the task, the skill runs. No pricing, no client data, no setup beyond that.
Open the crue · House Project. Mount the confidential layer from crue-house (GitHub connector, read-only, then Sync). Enable the Bexio connector here, House-only. This is where pricing, client data, offers and invoices live. Run offer, invoice, initialize-project and work-summary from here.
Every client's brain (brand-pack, imagery lock-ins, reference packs, decisions) lives in one private repo, crue-os/crue-clients, under clients/<slug>/. Each folder mounts into that client's Project as read-only knowledge. The per-client twin of crue-house.
lock-to-client skill picks the right file, applies the change in place, checks for conflicts against locked values, logs the decision.crue-clients/clients/<slug>/<file>.md through the GitHub write MCP. A clean message, never a force-push, never history.The GitHub token writes to crue-clients and crue-projects only. Never crue-os or crue-house. Pricing and proof are never writable from a client chat, and one client's folder is never written from another's Project.
In Claude Code it is automatic: the repo is cloned, the skill edits the file and commits and pushes. In a Project chat the write MCP is what makes it automatic. Same loop, two surfaces.
One client equals one Project equals one brand. Here is the whole setup.
initialize-project. Pick the client from Bexio or create them as a contact. It creates clients/<slug>/ in crue-clients from the latest template, filled, and commits it.Client · <Name>.crue-os/crue-clients → select only this client's folder clients/<slug>/ (read-only) → it syncs; Re-sync to refresh. The folder picker keeps every other client out, one project sees one client. Prereq: the Claude GitHub app has access to the private repo. Set the archetype in brand-pack. (Personal project with no GitHub in Files? Upload the five files instead.)work-summary (no rates), then run invoice from House to price it and create it in Bexio.clickup-brief tie a task back to its client: 💼 Client / Project (text), 🤖 Claude Project (URL to the Claude Project), 🧠 Github Client Brain (URL to the client folder in crue-clients). Set on the main task, returned as links.