Four steps from zero to a working connection.
Create a Personal Access Token
Open Dashboard → Settings → MCP. Click Create token, give it a label (the client you'll use it from), and pick scopes. Start narrow — you can revoke and re-create with different scopes at any time.
Shown Once
The raw token is displayed exactly once at creation. Copy it immediately; we store only a SHA256 hash and can't recover the original after the dialog closes.
Add the server to your MCP client
Every MCP-capable client uses the same JSON shape. For clients with a local config file, paste this block into the relevant file:
{
"mcpServers": {
"vaybel": {
"type": "http",
"url": "https://mcp.vaybel.com/",
"headers": {
"Authorization": "Bearer vbl_pat_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx"
}
}
}
}
Config locations for the common clients:
- Claude Desktop —
~/Library/Application Support/Claude/claude_desktop_config.json - Claude Code —
~/.claude.json - Cursor — add through
Settings → MCP → Add custom server - ChatGPT — add as a custom connector from
Settings → Connectors(paste the URL and Bearer header in the UI; no local file)
Restart the client after saving — MCP server configs are loaded on startup.
Keep The Token Out Of Git
A committed config with an embedded Bearer will be flagged by GitHub secret scanning (the vbl_pat_live_ prefix is registered). If you need to share a config file, move the PAT into an environment variable and reference it.
Verify the connection with ping
Ask your agent to call the ping tool, or send the JSON-RPC frame directly:
curl -N -X POST https://mcp.vaybel.com/ \
-H "Authorization: Bearer vbl_pat_live_..." \
-H "Content-Type: application/json" \
-H "Accept: application/json, text/event-stream" \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"ping","arguments":{"echo":"hello"}}}'
A successful response carries structuredContent.echo = "hello". A 401 means the Authorization header didn't land — re-check your client config and restart. Any other error shape: head to Tool Reference to match against the expected return.
Make a real call
From the MCP client, ask the agent something concrete:
List ten AOP hoodie blanks from the catalog.
The agent will call catalog.list_blanks with technique: "aop" and category: "hoodie". From there you can chain design.generate_design → design.wait_for_design → mockup.generate_mockup → mockup.get_mockup_status — the full design-to-preview loop without leaving the chat.
Want the full tool surface? See Tool Reference. Planning to run this in production? Read Authentication & Scopes first.