Frequently asked questions

The basics

What is Wakeshift?

Wakeshift lets an MCP server wake the AI agents that use it, so they hear back. When a long job finishes, a person replies or anything else happens in your app, your server calls emit, and the conversation that asked wakes up with it, in Claude Code, Codex, Cursor, Copilot CLI, a ChatGPT Work chat or another agent. It's a hosted service with SDKs for TypeScript, Python and Go, at wakeshift.ai.

Who is it for?

Companies that run an MCP server whose work finishes after the tool call returns: a render, a crawl, a build, a reply, a call. Your users' agents are the ones woken; your users don't need a Wakeshift account.

How do I add it to my MCP server?

Install the SDK (npm install @wakeshift/sdk, pip install wakeshift, or go get github.com/wakeshift/wakeshift-go), define your events, call attach on your server, and call emit when something happens. The quickstart takes you through it, and MCP frameworks lists the frameworks it works with, including Vercel's mcp-handler and Cloudflare's createMcpHandler. Other languages use the HTTP API. npx -y @wakeshift/cli init sets up your app's key from the terminal, and verify checks your server (the CLI). Using a coding agent? Paste this prompt: Add Wakeshift to this MCP server following https://wakeshift.ai/AGENTS.md.

Can I try it without signing up?

Yes. The try app gives you a URL that wakes your own agent: point curl, GitHub or Stripe at it. It takes about two minutes.

Agents

Which agents does it wake?

Claude Code, Codex, Cursor (CLI and IDE), GitHub Copilot CLI, opencode, Antigravity CLI, Claude Cowork (while it's open), OpenClaw (while its gateway runs) and Hermes Agent (while it runs) wake while idle, most within seconds. Grok Bot wakes while it's waiting for events. In ChatGPT's Work chats (ChatGPT Business or Enterprise) and dots, once the user asks ChatGPT to watch the event, each event starts a new run. Any other MCP client that can run a command in the background, and is told when it exits, is woken too. Cursor's cloud agents, Goose and Cline can't be woken from outside; their events arrive with the agent's next call. Agents has the table.

Does it wake the conversation the user has open, or start a new one?

The one that subscribed, in every agent that can be woken while idle. The exception is a client that subscribes to MCP Events itself, where each event starts a new run: ChatGPT (in Work chats and dots), and OpenClaw and Hermes Agent when they subscribe that way. Through Wakeshift's plugin, OpenClaw and Hermes wake the conversation instead.

Do I write code for each agent?

No. Wakeshift recognizes which agent subscribed and wakes it the way that agent can be woken. Your code is the same for all of them.

What do my users have to install or approve?

Claude Code, Cursor, Copilot CLI and Antigravity CLI need nothing installed. Codex, opencode, OpenClaw and Hermes Agent need a one-time install, which the agent offers and runs only if the user agrees, through each agent's own plugin command where it has one; Codex then asks the user to trust the new hook. Some agents ask before starting the background command that waits. Agents lists each step, and how to uninstall.

What happens if the laptop is closed or the session ends?

Wakeshift holds what arrives and delivers it once the conversation reconnects: by itself in Claude Code with Wakeshift's plugin, opencode and OpenClaw, and with the user's next message in Codex, Cursor, Copilot CLI and Antigravity CLI. In Claude Code without the plugin, your app's wait tool, if it's on, keeps waiting through /compact, closing and claude --continue (in the terminal, on accounts with MCP tasks); otherwise Claude Code starts waiting again when it's next asked to watch. Agents has each agent's steps.

Can an event make the agent do something on my users' behalf?

Events arrive marked as information from your app, never as messages from the user, so the agent can tell them from its user's words. It decides what to do, with whatever permissions its user already gave it.

Adding your app to their agent is the consent. A subscription starts only when the user asks their agent to watch something, and only if your tool allows it. Users can see and stop their subscriptions by asking their agent, and Wakeshift keeps an audit log of each consent and revocation.

Does Wakeshift answer the agent's permission prompts?

No. Those stay between the agent and its user.

Compared with other options

Can't I use MCP notifications?

Progress notifications travel inside a tool call that's still open. Once the call returns, nothing carries them. MCP's other notifications tell the client that a resource or a list changed; none of them is defined to start the agent's turn.

What about MCP Events (SEP-3415)?

It's the protocol's draft standard for events, and Wakeshift serves it: clients that speak it, such as ChatGPT in Work chats and dots, OpenClaw, Hermes Agent and Mastra, subscribe natively once your sign-in identifies the user. The local coding agents don't subscribe to other companies' events yet, so Wakeshift reaches them through its own adapters, from the same emit. Hosted MCP Events has the detail.

Which wakes an idle Claude Code session: MCP notifications, MCP tasks or Channels?

MCP tasks and Channels can; MCP's standard notifications can't. In Claude Code's terminal, on accounts where Anthropic has turned MCP tasks on, a finished task from a server that speaks protocol 2026-07-28 wakes the idle session. Channels push events into an open session, but they're a research preview: a channel plugin needs Anthropic's allowlist (or the user's organization's), Claude Code has to start with --channels, and a server on protocol 2026-07-28 isn't registered as a channel. MCP's standard notifications, such as progress or a changed resource, aren't defined to start a turn. Wakeshift doesn't need any of them: Claude Code asks the user once how to be notified (in this session, in every session through Wakeshift's plugin, or not at all) and wakes through the plugin or a background monitor, and the wait tool, which your app can turn on (it's off by default), uses MCP tasks where they're on.

How is it different from Hookdeck?

Hookdeck's MCP Events bridge turns third-party webhooks into MCP Events, and wakes Claude Code through a plugin. It's free, open source and verifies many webhook senders. Wakeshift is for MCP builders whose own events should reach every agent their users run, including Codex, Cursor and Copilot CLI, under per-user consent. The comparison has the detail.

Should I build MCP Events myself or use a hosted service?

If your users are only in ChatGPT's Work chats or dots, a free toolkit is enough. If they're in coding agents, the standard doesn't reach them yet, and the draft can join the spec no earlier than March 2027. With Wakeshift you integrate once: it speaks the standard to clients that subscribe and wakes the agents that don't. Hosted MCP Events has the detail.

Pricing and licensing

How much does it cost?

Free includes 50,000 deliveries a month. Starter is $29 a month with $0.10 per 1,000 deliveries beyond that, and Custom is for larger apps. A delivery is one event reaching one agent; retries never count. The people who use your app pay nothing. Pricing.

Is Wakeshift open source?

The SDKs, the connector, the agent plugins and the example apps are, under Apache 2.0, at github.com/wakeshift. The hosted service isn't.

Is there a SOC 2 report?

Not yet; it's in progress. Write to support@wakeshift.ai with your security questions.