MCP server

Let Claude, Cursor and other AI assistants build on Kerabie and work with your account.

Kerabie has a Model Context Protocol server. It does two jobs.

Build with Kerabie. Your coding assistant can search the docs, list API endpoints, read the TypeScript, React Native and Flutter SDK types, and follow integration guides, so it writes correct code for the API, the TypeScript SDK, the React Native SDK, the Flutter SDK, the web widget and WordPress. This works without an API key.
Work with your account. With an API key, the same assistant can use your live data. Connect it to Claude Desktop, Claude Code, Cursor or any MCP client and you can ask things like "which tickets are about to breach SLA?", "draft a Knowledge Base article from this conversation" or "pause the pricing-page campaign".

The account tools are a thin layer over the API and @kerabie/sdk, so they do exactly what your API key is allowed to do, and nothing else. The developer tools answer from documentation bundled with the package, so they match the SDK version you install.

Hosted server (no install)

Point your client at the hosted endpoint and send your API key as a Bearer token. Nothing runs on your machine.

https://mcp.kerabie.com/mcp
Authorization: Bearer ker_live_...

Claude Code:

claude mcp add --transport http kerabie https://mcp.kerabie.com/mcp --header "Authorization: Bearer ker_live_..."

Cursor (.cursor/mcp.json):

{
  "mcpServers": {
    "kerabie": {
      "url": "https://mcp.kerabie.com/mcp",
      "headers": { "Authorization": "Bearer ker_live_..." }
    }
  }
}

Add ?readOnly=true to the URL to leave out every tool that changes data. Without a key you get the developer tools only. Your key is used for every call and is never stored, so the server can do exactly what your key allows.

Get an API key

The developer tools work without a key. To let your assistant work with your account, create an API key under Settings → Integrations → API Keys (API access is included on every plan) and send it as the Bearer token shown above.

Careful: an MCP client can use every tool the key allows, including ones that reply to customers or delete things. Start with ?readOnly=true on the URL, and keep the client's approval prompts on for anything that changes data.

Developer tools

Available with or without an API key. None of them change anything.

Tool	What it does
`search_docs`	Search the docs, SDK readmes and API reference.
`get_doc`	Read a docs page, or one section of it.
`list_docs`	Every page with its description, by section.
`list_api_endpoints`	Every REST endpoint with its docs page and the equivalent TypeScript SDK call.
`get_sdk_types`	The TypeScript, React Native or Flutter SDK type declarations: list them, or fetch the ones you name.
`get_integration_guide`	The pages to follow for `typescript`, `react-native`, `flutter`, `wordpress`, `web-widget`, `rest-api`, `mcp` or `app-agent`.
`plan_agent_for_app`	The plan and checklist for building an AI agent for an app. See below.

The server also offers ready-made prompts (integrate_react_native, integrate_flutter, integrate_typescript_sdk, integrate_rest_api, install_web_widget, build_app_agent) and every docs page as a resource (kerabie-docs://sdks/react-native) for clients that support them.

Build an AI agent for your app

Ask your assistant to build the agent and it follows plan_agent_for_app, which returns an ordered plan for your platform (react-native, flutter or web-widget) and the things customers should be able to do (catalog, cart, addresses, scheduling, orders, payments, support). Leave the list out to get every checklist. The build_app_agent prompt starts the whole job.

The plan covers, in order:

What the app already has. It reads package.json and the native config first. If the app already has LiveKit it reuses it and checks the versions. If the app has another calling or WebRTC library (Zego, Agora, Twilio, Daily), it does not add LiveKit blindly: it builds the chat agent first and asks you before adding voice. See If your app already has LiveKit.
Platform requirements. The chat agent needs no native code and runs in Expo Go; voice needs a development build.
Rules for every tool. The customer comes only from X-Kerabie-User-Token, anything that changes something is secure, orders and bookings are confirmed on screen, addresses are never taken by voice, and lists are paginated.
The tools. Plain HTTP tools by default, an MCP server only when the app already has one or has many tools. A cart that lives on the device becomes app actions, not an invented server cart. Checklists for scheduling (slots, timezones, idempotency, rescheduling), addresses, orders, payments and support handoff.
Connect and choose routes. create_ai_tool with secure: true, or connect_mcp_server then update_mcp_server with secureTools.
Add the agent to the app, test the real flows, and report back what was built and what could not be.

The plan tells the assistant what to build and in what order. It does not write the code for you, and the assistant should say what it could not test.

Account tools

These need KERABIE_API_KEY.

Tools that change something are marked write; those that delete are also marked destructive, so clients can ask before running them.

Area	Tools
Conversations	`list_conversations`, `get_conversation`, `get_customer_history`, `take_over_conversation` (write)
Tickets	`list_tickets`, `get_ticket_stats`, `get_ticket`, `create_ticket` (write), `update_ticket` (write), `reply_to_ticket` (write)
Contacts	`list_contacts`, `get_contact_stats`, `get_contact`, `create_contact` (write), `update_contact` (write)
Knowledge Base	`search_knowledge_base`, `get_kb_stats`, `get_kb_insights`, `get_article`, `create_article` (write), `update_article` (write), `delete_article` (destructive)
Campaigns	`list_campaigns`, `get_campaign_stats`, `get_campaign`, `create_campaign`, `update_campaign`, `duplicate_campaign` (write), `delete_campaign` (destructive)
Analytics	`get_analytics_summary`, `get_analytics_insights`, `get_agent_stats`, `get_ticket_sla`, `get_setup_status`
Visitors and store	`list_visitors`, `get_visitors_overview`, `list_store_products`
AI tools	`list_ai_tools`, `get_ai_tool`, `create_ai_tool` (with `secure` for tools that act as the signed-in customer), `update_ai_tool`, `duplicate_ai_tool` (write), `delete_ai_tool` (destructive), `test_ai_tool` (write: calls your endpoint)
MCP servers	`list_mcp_servers`, `connect_mcp_server` (write), `sync_mcp_server` (write), `update_mcp_server` (write; `secureTools` sets which tools act as the signed-in customer), `test_mcp_server_tool` (write), `disconnect_mcp_server` (destructive)

Long results are cut at 60,000 characters with a note to use page, limit or filters. API errors come back as readable messages, for example Kerabie API error 403: Your plan includes 1 AI agent tool. Upgrade to add more.

Use MCP servers from your chatbot

The other direction: your AI chatbot can use tools from any remote MCP server. See Connecting MCP servers.