OpenAI has spread a complete automation stack across several ChatGPT features, and together they cover the same ground that platforms such as Zapier sell as one monthly subscription. ChatGPT Sites acts as an always-on server, MCP Events and app triggers start workflows, Scheduled Tasks carry out the work, and connected apps such as Gmail or Slack supply the hands that touch real data. The pieces arrived separately, which appears to be why few users have noticed that they form a single system. For a typical small business, the practical effect is that a form submission or inbound email can now set off a background task inside ChatGPT with no separate middleware subscription and no one at the keyboard. The main catch is that only a handful of third-party apps can send those wake-up events so far, and the runs are not instant.
The four parts of any automation
Every automation, regardless of the vendor, needs the same four components. A server listens around the clock. A trigger fires when something happens. A worker holds the instructions and executes them. And hands reach into connected applications to read and write data. Zapier-style platforms package all four into one product and charge for the bundle. OpenAI now offers each one as a separate feature inside ChatGPT.
| Component | What it does | ChatGPT feature |
|---|---|---|
| Server | Stays on and listens for inbound signals | ChatGPT Sites, hosted on Cloudflare Workers with a D1 database (10 GB in the documentation) and R2 object storage |
| Trigger | Starts the workflow | Native app triggers (Gmail, Slack, GitHub on eligible plans, introduced in August) and MCP Events (announced at DevDay on September 29, 2026) |
| Worker | Stores prompts and skills, runs when woken | Scheduled Tasks |
| Hands | Acts inside external apps | Connected app plugins: Gmail, Google Drive, Docs, Sheets, Calendar, Slack, GitHub |
ChatGPT Sites deserves a closer look because it is easy to mistake for a simple website builder for games or dashboards. It is a hosted server with persistent storage and a database, paid for by OpenAI, and OpenAI's documentation states that a plugin built with Plugin Creator can have its underlying MCP server hosted directly on a Site rather than on an outside platform such as Vercel or Supabase. Behind whatever interface a Site shows the public, it can receive webhooks, validate and store records, and signal ChatGPT.
From reactive to proactive: how MCP Events work
MCP, or Model Context Protocol, is the standard that lets ChatGPT call tools in connected apps. In ordinary tool calling, ChatGPT asks and the app answers: look up a lead, save a result. MCP Events reverse the direction. The app sends an inbound notification to ChatGPT announcing that something changed, such as a new form submission, and that notification wakes a Scheduled Task that has been waiting with its instructions already loaded.

▲ MCP Event waking a background task
A smartphone comparison makes the model easy to hold in mind. ChatGPT is the phone, plugins are installed apps, tool calls are the user opening an app to check something, and MCP Events are push notifications. The comparison also explains a common setup mistake: installing a plugin does not let it take initiative, just as installing a phone app does not grant it notification rights. The user must create an event subscription in ChatGPT, enable notifications for that plugin in the plugin settings, and tell the model which task to run when an alert arrives. ChatGPT's plugin settings expose these permissions along with tool risk levels, including a control to allow low-risk tools.
The result is a ChatGPT that works while the user sleeps or steps away, instead of one that waits for a typed or spoken prompt.
A lead-intake workflow, step by step
A working example shows how the pieces connect. A website offers a personalized AI starter guide to anyone who fills in a Google Form with their name, email, business role and the kind of help they want. From there the chain runs without human involvement:
- A prospect submits the Google Form.
- A small Google Apps Script picks up the submission and forwards the payload to a backend application hosted on a ChatGPT Site, which runs on a Cloudflare Worker.
- The backend saves the lead, validates the fields against business rules, and writes a log entry to its database.
- The backend sends an event notification through an existing subscription to ChatGPT.
- An event-based Scheduled Task wakes up, reads its stored instructions, and makes MCP tool calls against the Site's plugin to read the lead details and update records.
- ChatGPT uses the connected Gmail tools to send the tailored starter guide to the prospect.
One design point stands out: the Apps Script never talks to ChatGPT. It only talks to the Site, which handles the MCP handoff. Routing raw webhooks through a backend that validates and logs them, rather than straight into a chat prompt, keeps the system controlled and auditable.
An administrative dashboard for this flow recorded, as of a snapshot on November 2, 10 events received, 9 ChatGPT acknowledgments, 4 verified email sends and 0 failures. Logged time-to-send values ranged from 101.9 seconds to 402.0 seconds, and a full run of the email workflow took roughly 9 minutes. These are figures from a single setup, not a benchmark.
A second example uses a native trigger with no custom site at all. A Scheduled Task named for managing brand deals fires on every new email in Gmail, runs a saved prompt and skill, and applies labels such as Creator Hooks, Customers, Sales pitches and Sponsorship directly in the inbox.
Don't send everything to the model
Having ChatGPT available as a worker does not mean every request should reach it. A guiding rule from the same setups: never spend AI compute on work that a predictable rule can do. A website chat widget demonstrates this with three tiers.

▲ Three-tier routing before the model
| Tier | Trigger | Who answers | Cost |
|---|---|---|---|
| 1 | Visitor clicks a predefined common question | Local database, instant | No model call |
| 2 | Visitor types a question that matches a previously answered one | Stored answer, instant | No model call |
| 3 | Question matches nothing | ChatGPT, limited to approved answers, polite redirects, or holding for human follow-up | Model call, rate-capped |
The widget's safety configuration loads 24 approved FAQ topics and 285 supported wordings, which is why most questions never reach tier 3. It also enforces caps of 10 native events per hour and 40 per day, and a maximum of 60 callback attempts per day, so that a malicious or careless visitor cannot drain the AI usage limits. The same configuration states a zero-storage policy for IP and MAC addresses.
This hybrid pattern is arguably the most transferable idea here. Deterministic rules and cached answers handle volume; the model handles only the novel cases, and even then within a narrow mandate.
What still holds this back
The single biggest bottleneck is adoption. Shipmail, Bird and Twelfth.ai already publish native MCP Events, covering mail delivery triggers, communication webhooks, and monitoring of competitor prices and project changes. But the broad categories a small business relies on most, including forms, CRMs, calendars, e-commerce stores, social media, community software, payment processors and support desks, have not caught up. Until they do, a bridge such as Apps Script plus a Site backend is required, which is exactly the kind of glue work automation platforms exist to remove. Once an app supports MCP Events natively, the bridge disappears: connect the app to ChatGPT, turn on notifications, define the task.
Other constraints are worth knowing before replacing a paid tool:
- Runs are not instantaneous. The email workflow above took about 9 minutes end to end.
- Schedules hosted on ChatGPT Sites executed roughly once per hour in testing.
- MCP Events availability is documented for work chats on ChatGPT web, the desktop app with Cloud selected, and Dots.
- The rollout is fragmented across Sites, app triggers, MCP Events, Scheduled Tasks and plugins, and the Cloudflare Worker layer is abstracted away, so the relationship between a Site and its plugin can be confusing at first.
Advanced automation professionals and developers will likely keep dedicated enterprise tooling. Workflows that need sub-minute latency, high-frequency schedules, or connectors to apps without event support still appear to be the territory of standalone platforms and the agencies that build on them.
What to take from this
OpenAI has assembled server, trigger, worker and hands inside ChatGPT, and for common small-business flows such as inbox triage, form-driven follow-up and website chat, that may remove the reason to pay a separate middleware subscription. The gaps are real for now: few apps emit MCP Events, cycles take minutes, and hourly schedules limit what Sites can do on their own.
If you already pay for ChatGPT and for an automation tool, a sensible path is to inventory your workflows and sort them. Flows that start from Gmail, Slack or GitHub can move to native triggers and Scheduled Tasks today. Flows that start from a form or another unsupported app can run through a Site backend if you are willing to maintain a small courier script, or stay where they are until the provider adds events. In every case, put rules and stored answers in front of the model, cap the rate at which events can reach it, and store instructions as saved prompts and skills rather than ad hoc chat messages. Keep the paid platform for whatever remains latency-sensitive or unsupported, and revisit the split as more apps adopt MCP Events.


