MCP server bridging GitHub webhooks via Cloudflare Worker for real-time event streaming
API key
WEBHOOK_WORKER_URL, WEBHOOK_CHANNEL
ตรวจสอบล่าสุด
2026-11-09
ดาว
—
เครื่องมือ
—
อัปไทม์
—
Cloudflare Worker receives GitHub webhooks, verifies signatures, stores events in a Durable Object with SQLite.
Local MCP bridge (.mcpb) proxies tool calls to the Worker and optionally connects via WebSocket for real-time channel notifications.
No local webhook receiver or tunnel required.
Breaking change: MCP protocol revision 2026-07-28
From this release the Worker serves MCP protocol revision 2026-07-28 only. It keeps no compatibility lane for the previous revision.
Bridge versions older than this release stop working. They open a session with initialize, which the Worker no longer answers. The failure is quiet: the bridge does not crash, it returns the protocol error as tool output text.
Real-time channel notifications keep arriving, which hides the breakage. The /events stream is not MCP and is unaffected, so a stale bridge still pushes event summaries while every tool call — including mark_processed — fails. The pending queue stops being cleared even though notifications look healthy.
Restart the MCP client to pick up the new bridge. The bridge is launched with npx, and @latest is resolved at process start — an already-running Claude Desktop, Claude Code, or Codex keeps the copy it started with, however new the published version is. Quit it fully and reopen.
Pinning the bridge version leaves you stuck. If your MCP client config pins a version older than this release, restarting does not help; remove the pin (or move it forward) first.
The Worker and the bridge ship together, so a bridge from this release or later needs no configuration change.
Prerequisites
Component
Required
Node.js 18+
MCP server
Cloudflare account
Worker deployment (self-hosting)
Getting Started
1. Install the GitHub App
Install the GitHub Webhook MCP app on your GitHub organization or account:
Choose which repositories to grant access to (or all repositories)
Approve the requested permissions
Note: When the app requests new permissions after an update, you must approve them in your GitHub notification or the app's installation settings. Webhooks will not be delivered until permissions are accepted.
Important: Do not create a separate repository webhook for the same endpoint. The GitHub App handles all webhook delivery — a repository webhook would cause duplicate or malformed requests.
2. Set up the MCP client
Continue to the Installation guide to connect your AI assistant to the webhook service.
MCP Client Setup for Claude Desktop, Claude Code CLI, and Codex
Self-Hosting Guide for Cloudflare Workers deployment
Updating
A published release does not reach a running client on its own. npx resolves the package version
once, when the process starts — including when the client config pins @latest — so an MCP client
that is already running keeps the version it started with no matter what the registry serves.
Restart the MCP client (Claude Desktop, Claude Code, Codex) to pick up a new release. The
restart is what moves the client onto the new version.
Check what the registry actually has with --prefer-online. The npm CLI caches registry metadata,
so a bare npm view can still report the previous version shortly after a publish:
npm view github-webhook-mcp version --prefer-online
Usage Examples
Example 1: Check pending webhook status
User prompt:
"Are there any new GitHub notifications?"
Expected output:
The AI calls get_pending_status and returns a summary:
You have 3 pending webhook events:
- 2 push events
- 1 pull_request event
Example 2: Inspect a specific event
User prompt:
"Show me the details of the latest pull request event."
Expected output:
The AI calls list_pending_events to find the PR event, then get_event with the event ID to retrieve the full payload:
PR #42 "Fix login timeout" was opened by @alice in repo acme/web-app
Branch: fix/login-timeout → main
Status: open
Changed files: 3
Example 3: Process events after review
User prompt:
"I've reviewed all the push notifications, mark them as done."
Expected output:
The AI calls list_pending_events to find push events, then clears them with a single
mark_processed({ event_ids: [...] }) call:
Marked 2 push events as processed:
- Push to main by @bob (3 commits)
- Push to develop by @alice (1 commit)
Example 4: Monitor CI status via webhooks
User prompt:
"Did the CI checks pass on my latest PR?"
Expected output:
The AI calls list_pending_events to find check_run events related to the PR, then get_event for details:
CI results for PR #42 "Fix login timeout":
- build (ubuntu-latest): ✓ passed
- lint: ✓ passed
- test (node-18): ✓ passed
All checks passed.
MCP Tools
Tool
Description
get_pending_status
Lightweight snapshot of pending event counts by type
list_pending_events
Summaries of pending events (no full payloads)
get_event
Full payload for a single event by ID
get_webhook_events
Full payloads for all pending events
mark_processed
Mark events as processed (event_id for one, event_ids for a batch)
Event Retention
Stored events are purged automatically to bound Durable Object storage. The Worker
runs a time-based sweep on a Durable Object Alarm (daily), so cleanup happens even
for tenants that never call mark_processed:
Event class
Retention window
Env var
Default
Processed (mark_processed called)
older than the window is deleted
PURGE_AFTER_DAYS
3 days
Unprocessed (never marked)
older than the window is deleted
UNPROCESSED_PURGE_AFTER_DAYS
90 days
The longer window for unprocessed events is intentional: unprocessed means
user-unseen, so the safety margin before dropping is wide (the 3-day vs 90-day
asymmetry is by design).
The sweep runs via a Durable Object Alarm on a daily cadence and reschedules
itself, so it fires independently of consumption. Processed events are also
purged immediately on mark_processed for promptness; the Alarm sweep is the
guarantee that covers abandoned tenants.
Both windows are configurable in worker/wrangler.toml ([vars]). Setting a
value to 0 purges that class immediately on sweep.
Known limitation: the windows bound event age, not volume. A high-rate,
never-consumed tenant can still reach Cloudflare's 1 GB-per-DO ceiling before
the 90-day window applies. A volume-based hard cap is tracked separately.
Monorepo Structure
worker/ — Cloudflare Worker + Durable Objects
local-mcp/ — Local stdio MCP bridge (TypeScript, dev)
mcp-server/ — .mcpb package for Claude Desktop
shared/ — Shared types and utilities
Privacy Policy
Events are stored in a Cloudflare Durable Object (edge storage). The local MCP bridge proxies tool calls to the Worker and does not store event data locally.
