Wakeshift vs Hookdeck, MCP Events toolkits and built-in wakes
Four ways to tell an agent that something it's waiting on has happened. They solve different problems, and for some cases Wakeshift isn't the right one.
The short answer
Wakeshift is the right choice when you run an MCP server whose jobs finish later, and your users' agents are Claude Code, Codex, Cursor, Copilot CLI and others. One emit wakes the conversation that asked, in whichever agent it is.
Something else may fit better when:
- You want GitHub, Stripe or Resend webhooks in your own Claude Code session: Hookdeck's MCP Events bridge is free, open source, and verifies many more webhook senders than Wakeshift does.
- Your users work only in ChatGPT's Work chats or dots, or in OpenClaw or Hermes automations: a free MCP Events toolkit is enough. Those clients subscribe themselves.
- The wait is a pull request or CI run inside one agent: use that agent's own feature. Claude Code's Projects, GitHub's Agent Merge, Cursor's Agent SDK and Devin each follow their own pull requests now.
Side by side
| Wakeshift | Hookdeck's MCP Events bridge | MCP Events toolkits | Each agent's built-in wakes | |
|---|---|---|---|---|
| What it is | A service and SDK for MCP servers: your events wake your users' agents | An open-source MCP server that turns third-party webhooks into MCP Events, with a Claude Code plugin | Server-side libraries for the MCP Events draft (Upstash, Vercel's mcp-handler, Alpic's Skybridge, Hookdeck Outpost) |
Features inside each agent: MCP tasks, Channels, codex queue, Cursor and Copilot automations |
| Who it's for | MCP builders | Developers wiring webhooks into their own agent | MCP builders | Each agent's own users |
| Agents it reaches | Every client in the next two columns, plus the conversations users already have open in Claude Code, Codex, Cursor, Copilot CLI, opencode, Antigravity CLI, Claude Cowork, OpenClaw, Hermes Agent and more | The clients that subscribe to MCP Events, plus Claude Code through its plugin | Clients that subscribe: ChatGPT's Work chats and dots, OpenClaw, Hermes Agent, Mastra, TanStack AI, the Vercel AI SDK | Only that agent |
| What wakes | The conversation that subscribed, while it's idle. In a ChatGPT Work chat, a new run starts in the chat | A new run in the clients that subscribe; in Claude Code, a running session between prompts | A new run, or your own code, in each client (Mastra's signal provider adds it to the subscribing thread) | Varies: an idle Claude Code terminal session (MCP tasks), a new cloud session (automations), or a message as if the user typed it (codex queue) |
| Events from | Your own code (emit), or any webhook through a publish URL |
Third-party webhooks | Your own code | GitHub, Slack, schedules, and each agent's own triggers |
| Consent | Adding your app is the consent; any install runs only if the user agrees; consents and revocations kept in an audit log | Each client's own | Each client's own | Each agent's own |
| Standard | Serves MCP Events (SEP-3415) to clients that subscribe, and its own adapters to agents that don't | MCP Events | MCP Events | Mostly each agent's own; Claude Code's MCP tasks follow the MCP Tasks extension |
| Price | Free up to 50,000 deliveries a month; Starter $29; Custom | Free (MIT); runs on a Hookdeck account, free tier or paid | Free (Vercel bills the receiving side at $0.95 per 1,000 deliveries on Pro) | Included in each agent's plans (some only on paid ones) |
| Open source | The SDKs, connector and plugins (Apache 2.0); the service isn't | Yes (MIT) | Yes, mostly | Codex's CLI, with codex queue, is (Apache 2.0); the rest aren't |
Hookdeck's MCP Events bridge
What it does well. Hookdeck's bridge (launch post) turns webhooks from GitHub, Resend, Stripe, Slack and any signing sender into MCP Events, on Hookdeck's mature event gateway, with retries, deduplication, replay and observability. It's MIT-licensed and free, and Hookdeck verifies signatures for many more webhook senders than Wakeshift's publish URLs, which can require Standard Webhooks signatures (Svix, OpenAI, Clerk and others). For a developer who wants third-party webhooks in their own agent, it's a strong choice.
Where it stops. Beyond the clients that subscribe to MCP Events, it reaches one coding agent: Claude Code, through a plugin. It doesn't wake Codex, Cursor, Copilot CLI, opencode or Antigravity CLI, and it has no per-user consent or sessions for an MCP builder's users: each bridge has one owner (mcp-events-bridge). Hookdeck says it's building subscriptions into Event Gateway, so clients can subscribe without running a bridge.
How Wakeshift relates. Different buyer, mostly. Hookdeck brings other companies' webhooks to one developer's agent. Wakeshift lets an MCP builder's own events reach every one of its users' agents, under consent each user gives once.
MCP Events toolkits
Free toolkits implement the server side of MCP Events (SEP-3415), the protocol's draft standard for events. If your users are only in ChatGPT's Work chats or dots, a toolkit and a few handlers are enough (OpenAI's MCP Events guide). They reach only the clients that subscribe, which start a new run for each event; the main local coding agents don't subscribe, and some toolkits leave the subscription store and delivery to your code. Wakeshift serves the same standard and wakes the coding agents too, from the same emit: hosted MCP Events has the detail.
Each agent's built-in wakes
What they do well. They're built in and come with the agent's plans, and inside one agent they're often the best answer:
- Claude Code wakes an idle terminal session when an MCP task finishes, on accounts where Anthropic has turned MCP tasks on and from servers that speak protocol 2026-07-28. Its Channels push events into an open session, as a research preview behind Anthropic's allowlist (Claude Code docs). Its scheduled tasks and Projects start or wake Claude's cloud sessions on GitHub events.
- Codex queues a message into an idle thread with
codex queue, and its automations start tasks from Gmail, Slack and GitHub events. - Cursor's automations start its cloud agents from GitHub, Slack, webhook and other events, and its Agent SDK wakes agents developers build, resuming a session on a follow-up to a pull request.
- GitHub Copilot's automations start Copilot's cloud agent from GitHub events and a REST trigger, and Agent Merge follows a pull request through failing checks.
Where they stop. Each wakes only its own agent. None takes events from another company's MCP server into whichever agent the user runs. Some deliver in ways an MCP builder can't use: codex queue delivers as if the user typed it, and Channels blocks third-party plugins outside the allowlist (claude-code#82571).
How Wakeshift relates. Wakeshift uses each agent's own mechanisms where they fit, such as Claude Code's MCP tasks for its wait tool, and adds the agents each built-in wake doesn't cover. It never delivers as the user.
Webhook and workflow platforms
Svix, Composio, Inngest and Trigger.dev often come up as ways to handle events. They're complements, not alternatives: each gets an event to your code, and none wakes an agent by itself.
- Svix sends and receives webhooks reliably. A publish URL accepts posts signed by a Svix sender, so a Svix webhook whose body matches your event can wake your users' agents without a handler of your own (publish URLs).
- Composio's triggers deliver an app's events to one webhook URL you control, not to an agent. Call
emitfrom that handler. - Inngest and Trigger.dev run durable jobs and workflows, and Inngest can pause a run until a matching event arrives. When the job a tool started finishes, call
emitfrom its last step.
What Wakeshift doesn't do
- Work with no prompts at all. Some agents need a one-time install, and some ask before starting the command that waits; in Claude Code without its plugin, manual permission mode asks again every 25 minutes, unless your app turns on the wait tool. Agents lists every step.
- Verify every webhook sender. Publish URLs can require Standard Webhooks signatures; the try app also checks GitHub's, Stripe's and Linear's. For other providers, verify in your own handler and call
emit. - Run agents or continue work on its own. It wakes the conversation that asked; the agent decides what to do. It doesn't move a session from one agent to another.
- Run on your own servers. Wakeshift is a hosted service.
Next
- Try it in two minutes, no sign-up.
- Long jobs: tell the agent when the render, crawl or build is done.
- One MCP server, every agent: what each agent asks of your user.
- FAQ