An MCP server that gives AI agents eyes on webhooks, inspect events, replay deliveries, mock signed payloads, create apps. Designed for the 45-second AI handler workflow.

Apache 2.0 · npm i -g anyhook-mcp

Zero-config (new in 0.2.0)

No API key needed to start:

claude mcp add anyhook -- npx -y anyhook-mcp

Then ask your agent to run anyhook_quickstart, it creates a free relay endpoint + API key instantly (no signup), and this MCP session auto-connects. All account tools (events, replay, apps) work immediately. The response includes a claim_url to keep the endpoint permanently.

What it does

Drop this MCP server into Claude Desktop / Cursor / Claude Code and your agent can:

anyhook_apps_list, see every app in your AnyHook account with its inbound URL
anyhook_apps_create, spin up a new app from a prompt
anyhook_events, list recent events, filter by app / status
anyhook_inspect, pull a single event's headers, body, signature status
anyhook_replay, re-send a stored event to its destinations (does not burn quota)
anyhook_undelivered, list events that never reached a destination
anyhook_replay_failed, bulk-replay every failed event for an app
anyhook_mock, generate a signed Stripe / GitHub / Slack payload for local handler testing
anyhook_verify, verify any incoming signature against your secret
anyhook_providers, list every webhook provider AnyHook supports

The inbound URL an app gives you accepts webhook POSTs and also answers provider setup handshakes automatically: Meta's GET carrying hub.mode=subscribe gets its hub.challenge echoed back, even before the app is configured, so the URL can be registered in the Meta App Dashboard first and wired up after.

Two modes:

Mode	When	What works
remote	`ANYHOOK_API_KEY` is set	All tools above, live against your AnyHook account
local	No API key	Provider toolkit (`mock` / `verify` / `providers`) + an in-memory store you can simulate events into. Useful for trying it out before signing up.

Quick start

npm install -g anyhook-mcp
# or run without installing
npx -y anyhook-mcp

Get an API key from https://anyhook.net/dashboard/settings/api-keys. Keys start with ahk_live_.

Install in Claude Desktop

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

{
  "mcpServers": {
    "anyhook": {
      "command": "npx",
      "args": ["-y", "anyhook-mcp"],
      "env": {
        "ANYHOOK_API_KEY": "ahk_live_xxxxxxxx"
      }
    }
  }
}

Restart Claude Desktop. The tools appear under the hammer icon.

Install in Cursor

Open Settings → MCP → Add new MCP server. Or edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "anyhook": {
      "command": "npx",
      "args": ["-y", "anyhook-mcp"],
      "env": {
        "ANYHOOK_API_KEY": "ahk_live_xxxxxxxx"
      }
    }
  }
}

Install in Claude Code

claude mcp add anyhook -e ANYHOOK_API_KEY=ahk_live_xxxxxxxx -- npx -y anyhook-mcp

Or edit ~/.claude/mcp.json directly with the same structure as above.

Remote endpoint (no local process)

The same server is hosted at https://anyhook.net/mcp (streamable HTTP, stateless). Use it from claude.ai custom connectors, ChatGPT, or any client that speaks HTTP, nothing to install:

{ "mcpServers": { "anyhook": { "url": "https://anyhook.net/mcp" } } }

Auth per request: send Authorization: Bearer ahk_live_..., or connect keyless and call anyhook_quickstart, it returns an api_key you then pass as an argument on account tool calls (the transport is stateless, so the session can't hold it for you).

Local-only mode (no account needed)

Skip the env var:

{
  "mcpServers": {
    "anyhook": {
      "command": "npx",
      "args": ["-y", "anyhook-mcp"]
    }
  }
}

The agent gets anyhook_mock, anyhook_verify, anyhook_providers, anyhook_simulate, anyhook_events (against memory store), and anyhook_inspect (against memory store). Useful for "generate me a signed Stripe payment_intent.succeeded and POST it to localhost:3000" flows during local dev.

Example conversations

Triage a failing endpoint:

"Show me undelivered events for the stripe-prod app from the last hour, then replay them."

The agent calls anyhook_undelivered then anyhook_replay for each, surfacing the responses inline.

Create an app for a new integration:

"Make a new AnyHook app called replicate-video-jobs, source generic, point it at https://my-app.vercel.app/api/replicate-callback."

Calls anyhook_apps_create, returns the inbound URL you paste into Replicate.

Sanity-check a signature failure:

"Here's an inbound payload from Stripe that AnyHook is rejecting. Verify the signature manually with secret whsec_xxx."

Calls anyhook_verify with the supplied headers + body + secret, returns valid: false plus the failure reason.

Mock a webhook for local testing:

"Generate a Stripe payment_intent.succeeded event and POST it to http://localhost:3000/api/webhooks/stripe."

Calls anyhook_mock with targetUrl set; returns the request that was sent + your handler's response.

Environment variables

Var	Required	Default	Purpose
`ANYHOOK_API_KEY`	for remote mode	-	`ahk_live_*` from https://anyhook.net/dashboard/settings/api-keys
`ANYHOOK_API_BASE`	no	`https://anyhook.net`	Override for self-hosted AnyHook deployments

Why this is interesting

AI agents have been blind to their own webhook infrastructure. They can write the integration code, but once an event misbehaves in production they can't see it, you tell them what happened. With this MCP server, Claude / Cursor can read the actual event log, replay deliveries, debug signature failures, and even spin up new apps. It closes the loop between writing webhook code and operating it.