Knock MCP Server

Knock's MCP server lets AI coding assistants manage your notification infrastructure — workflows, channels, templates, users, and more — directly from tools like Cursor, Claude Code, and Claude Desktop.

This remote MCP server acts as middleware to the Knock API, authenticated via Knock's OAuth flow and optimized for developer workflows.

Getting Started

Connect your AI assistant to Knock's MCP server in seconds. No local setup required.

Cursor

Add the following to your Cursor MCP configuration (~/.cursor/mcp.json):

{
  "mcpServers": {
    "knock": {
      "type": "http",
      "url": "https://mcp.knock.app/mcp"
    }
  }
}

Claude Desktop

Add to your Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "knock": {
      "type": "http",
      "url": "https://mcp.knock.app/mcp"
    }
  }
}

On first connection, your browser will open to authorize and select which capabilities to grant.

Related MCP server: ClickUp MCP Server

Capabilities

When connecting, you choose exactly which tool groups to enable. By default, only Management API (code mode) is on; the classic per-resource tools are opt-in.

Authentication

The MCP server uses OAuth 2.1 + PKCE via Knock's AuthKit. When you first connect, you'll be directed to authorize the connection and select which capabilities to grant. Your credentials are never stored by the MCP server — it exchanges tokens with Knock's API on your behalf.

Self-Hosting & Local Development

If you need to run the MCP server yourself (e.g. for development or custom deployments), read on.

Prerequisites

• Node.js 18+

• A Cloudflare account with Workers enabled

• A Knock account with AuthKit configured

1. Install dependencies

npm install
npm install --prefix client

The root .npmrc sets legacy-peer-deps=true because agents declares optional peer packages (for example ai) that this Worker does not use. @modelcontextprotocol/sdk is pinned to the same version as agents so TypeScript sees a single McpServer type.

2. Create a KV namespace

wrangler kv namespace create OAUTH_KV

Copy the returned id into wrangler.jsonc:

"kv_namespaces": [
  {
    "binding": "OAUTH_KV",
    "id": "your-namespace-id-here"
  }
]

3. Configure environment variables

cp .dev.vars.example .dev.vars

Dynamic Worker loader (Code Mode): this Worker declares a worker_loaders binding named LOADER in wrangler.jsonc for @cloudflare/codemode. Use a current compatibility_date and a recent wrangler / Workers runtime.

Production URLs are set in wrangler.jsonc under vars (KNOCK_AUTH_URL, KNOCK_DASHBOARD_URL). Override them with secrets only if you need per-environment values.

COOKIE_ENCRYPTION_KEY must be a Wrangler secret — do not add it to vars. Cloudflare rejects the same binding name as both a var and a secret.

wrangler secret put COOKIE_ENCRYPTION_KEY

SENTRY_DSN should be a Wrangler secret in any environment where errors are reported:

wrangler secret put SENTRY_DSN

Use a random 32-byte hex value (e.g. openssl rand -hex 32). After changing config, deploy so the Worker no longer declares a conflicting var.

The generated worker-configuration.d.ts (from wrangler types) types Env including secrets such as COOKIE_ENCRYPTION_KEY and optional .dev.vars entries. See src/env.d.ts for notes. You do not need a production DEV_ORIGIN unless you use the same origin-rewrite pattern as local dev.

4. Run locally

npm run build:client
npm run dev

The worker serves MCP at http://localhost:8788/mcp. Ensure .dev.vars sets DEV_ORIGIN=http://localhost:8788 so OAuth redirects and metadata match your machine.

4b. Debug with MCP Inspector

The repo includes mcp-inspector.json, which points the MCP Inspector at your local Streamable HTTP endpoint.

1. Terminal A — run the worker (after a client build, if assets changed):

   npm run build:client
   npm run dev

2. Terminal B — start the inspector (installs via devDependency on first npm install):

   npm run inspector

3. Open http://localhost:6274 in your browser. The UI should open with Streamable HTTP and http://localhost:8788/mcp already selected (from the config).

4. Click Connect. Complete the Knock OAuth and tool-selection flow in the browser when prompted.

The inspector also runs an MCP proxy on http://localhost:6277 by default. Override ports if needed, for example:

CLIENT_PORT=8080 SERVER_PORT=9000 npm run inspector

Note: @modelcontextprotocol/inspector currently recommends Node.js 22.7.5+; use a recent Node version for the inspector UI.

5. Deploy

npm run deploy

This builds the client UI and deploys to Cloudflare Workers. Update wrangler.jsonc with your custom domain:

"routes": [
  {
    "pattern": "mcp.your-domain.com",
    "custom_domain": true
  }
]

Architecture

MCP Client (e.g. Cursor, Claude Desktop)
    │
    ▼ OAuth 2.1 + PKCE
Cloudflare Worker (this repo)
    │  ├─ /mcp          — MCP endpoint (Durable Object)
    │  ├─ /authorize    — OAuth consent + tool selection UI
    │  └─ /callback     — Token exchange with Knock AuthKit
    │       search_mapi / execute_mapi  — Code Mode (OpenAPI + Dynamic Worker)
    │       optional classic toolkit tools
    │
    ▼
Knock Management API (control.knock.app) and other Knock APIs

The worker is deployed on Cloudflare Workers with Durable Objects for stateful MCP sessions. It acts as both the OAuth authorization server (to MCP clients) and an OAuth client (to Knock's AuthKit). Dynamic client registration means the worker registers itself with AuthKit at runtime — no static client IDs or pre-registered redirect URIs needed. Code Mode uses @cloudflare/codemode with a Worker Loader so LLM-generated code runs in an isolated sub-worker; mapi.request() on the host applies your OAuth token to https://control.knock.app.

License

MIT — see LICENSE.