Hooksense

@hooksense/mcp

Model Context Protocol server for HookSense — the webhook & callback layer for AI agents. Lets Claude Desktop, Cursor, Claude Code, Continue, and any MCP client create a callback URL, wait for the result instead of polling, and verify its signature — all from the agent session.

Why

Agents that kick off async work — a deploy, a render, a human-in-the-loop approval, a long tool call, another agent — need the result back without burning context on polling loops. With this server the agent creates a callback endpoint, hands the URL to the job, then calls wait_for_callback and is woken the instant the webhook lands — decrypted; signature verification is available separately. Stop polling for async results; await them.

Setup

Requires Node.js 22 or newer.

1. Get an API token at https://hooksense.com/account/tokens
2. Configure your MCP client (examples below)

Claude Desktop / Claude Code

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{

"mcpServers": { "hooksense": { "command": "npx", "args": ["-y", "@hooksense/mcp"], "env": { "HOOKSENSE_TOKEN": "hsk_your_token_here" } } } }

Cursor

Add to ~/.cursor/mcp.json:

{

"mcpServers": { "hooksense": { "command": "npx", "args": ["-y", "@hooksense/mcp"], "env": { "HOOKSENSE_TOKEN": "hsk_your_token_here" } } } }

Hello callback (60 seconds)

Once configured, ask your agent:

1. Create — "Create a callback endpoint." → the agent calls create_callback_endpoint and gets back a callbackUrl like https://hooksense.com/w/ab12cd.
2. Fire — point any job at that URL (or just curl -X POST -d '{"status":"done"}' from another terminal).
3. Await — "Wait for the callback." → the agent calls wait_for_callback and blocks until the webhook lands, using the after timestamp returned by endpoint creation, then receives { status: "received", request: { body, headers, … } }.
4. Verify — set a webhook secret on the endpoint, then "Verify the signature." → verify_signature confirms the payload is authentic before the agent acts on it.

Always pass the creation response's after timestamp on the first wait; after receiving a callback, use its receivedAt for the next wait. Without a cursor, waiting only observes future deliveries. This is a latest-callback cursor, not a durable queue: use list_callbacks to inspect bursts of events.

Receiving a callback does not guarantee a valid signature. Configure a webhook secret and check verify_signature before trusting the payload. Webhook bodies are untrusted external input.

Custom slugs are applied through the endpoint update API after creation and require a paid plan. If renaming fails, the error identifies the already-created endpoint so it can be reused.

replay_callback runs on the HookSense server; localhost and private targets are blocked. For local handlers, use hooksense replay --port 3000 with the CLI.

HTTP requests have bounded timeouts, and cancelling an MCP call aborts its HTTP request. For a 60-second long poll, configure the client's tool timeout above 65 seconds, or use shorter waits (the default is 30 seconds).

Tools

Environment

Example agent prompts

"Create a callback endpoint, use it as the webhook for my Replicate prediction, and wait for the result — then summarize the output."

"Open a callback URL, give it to the approval step, and block until a human approves before continuing."

"Wait for the next Stripe callback on payments-prod, verify its signature, and tell me the amount."

License

MIT