VibeCheck

The CMS you talk to. A git-backed, schema-safe, headless CMS you edit by chatting with an AI agent.

🔗 Live demo — itself built with VibeCheck. You bring the vibe — a sentence to the agent — it does the check: validates against a schema, commits to your repo, and your host redeploys. Content is typed data, not markup, so an agent can't break the build.

Part of the EpleCheck family. Built with Astro + the Model Context Protocol.

You: "Add a pricing page with three tiers."
→ agent writes structured content → commits to GitHub → Vercel ships it. ~30s.

Why

• Edit by conversation — no admin UI, no dashboard. Just talk to Claude.
• Can't break the build — every change is schema-validated. A bad edit fails the check, not the site.
• Git-native — every edit is a reviewable commit / pull request. Full history, rollbacks, the workflow you already trust.
• Deploys itself — static output to Vercel / Netlify / Cloudflare Pages.

Layout (npm workspaces monorepo)

packages/schema   @vibecheck/schema  — the content contract (zod). One source of truth.
apps/site         the Astro site     — renderer + your content. This is what deploys.
apps/mcp          @vibecheck/mcp     — the MCP server: lets an agent edit from anywhere.

site and mcp both import @vibecheck/schema, so the contract can never drift.

Quickstart

1. Use this template → your own GitHub repo (or git clone).
2. Deploy to Vercel — import the repo. For this monorepo, set the project Root Directory to apps/site. (Netlify / Cloudflare Pages / GitHub Pages work too — it's static output.)
3. Edit by chat, two ways:
   • Claude Code (local, zero infra): open the repo in Claude Code and ask it to edit pages. It writes the YAML directly; the schema validates at build.
   • MCP server (edit from anywhere): run the VibeCheck MCP and add it to Claude as a connector — edit from claude.ai, Desktop, or your phone:

     cp apps/mcp/.env.example apps/mcp/.env   # fill in GITHUB_TOKEN + GITHUB_REPO
     npm run build && npm start -w @vibecheck/mcp
     

4. Talk to it: "Create an FAQ page", "change the hero headline", "unpublish the pricing page" → committed → live.

Content model

A page is a YAML file in apps/site/src/content/pages/**/*.yaml — a title and a list of typed sections:

Section	What it is
hero	heading, subheading, CTA
features	grid of cards (optional links / images)
pricing	plans with features + CTAs
cta	a call-to-action band
richtext	the only section with raw HTML
form	a simple form (name/email/fields)

Only richtext carries HTML, so machine-generated content stays safe. The schema lives in packages/schema/src/index.ts.

The MCP server

@vibecheck/mcp exposes tools (create_page, update_page, get_page, delete_page, rename_page, set_draft, list_pages, plus blog-post tools) that commit content into your repo. Configure it via apps/mcp/.env:

• GitHub (default): GITHUB_TOKEN, GITHUB_REPO (owner/name).
• GitLab (optional): set VIBECHECK_PROVIDER=gitlab + GITLAB_*.
• MCP_AUTH_TOKEN — clients send this as the Bearer token.

Publish modes on every write tool: merge_request (default — opens a PR for review) or direct (commit to the default branch; your host deploys on push).

Run it — locally (npm start -w @vibecheck/mcp), or with Docker:

cp apps/mcp/.env.example apps/mcp/.env   # fill in token + repo
docker compose up -d --build             # or pull the prebuilt image:
# docker pull ghcr.io/eplecheck/vibecheck-mcp:latest

It listens on :8787 (POST /mcp); add it to Claude as an MCP connector with Authorization: Bearer <MCP_AUTH_TOKEN>.

CI

• build.yml runs npm run build on every push/PR — because the site build validates every page against the schema, a green build means all content is valid.
• docker.yml builds the MCP image and publishes it to ghcr.io/eplecheck/vibecheck-mcp.

License

MIT © EpleCheck AS.