ChatGPT now gives you three distinct ways to hand off recurring work. Pages are living documents that can refresh themselves on a schedule. The OpenAI Developer Platform lets you design agents, AI programs that take a goal and carry out several steps on their own. And Dot is a personal agent that runs on its own virtual computer and keeps working after you close the chat. All three become far more useful once they can reach your other apps, and for apps that ChatGPT’s built-in plugins do not cover, Zapier’s MCP server fills the gap. MCP, or Model Context Protocol, is a standard way for an AI model to call outside tools. Below is what each path is good for, how to set it up, and what to check before you let it run unattended.
The three paths at a glance
| Path | Where you build it | Best suited for |
|---|---|---|
| Pages | Spaces inside ChatGPT | Dashboards that refresh daily, shared team documents |
| Platform agents | OpenAI Developer Platform | Repeatable jobs with a fixed output format |
| Dot | ChatGPT’s personal agent | Multi-app tasks that should finish while you are away |
Pages: documents that update themselves
Pages live in Spaces, found in ChatGPT’s left sidebar. Spaces is a hub that gathers your videos, images, personal library and pages in one place. A page is an interactive canvas where you and your team write, build and collaborate with ChatGPT. A schedule manager in the upper-right corner lets you attach recurring background jobs to a page, and comments and a share menu let you invite teammates, outside users or AI assistants with set permissions.
Most editing happens through short commands:
- Type a forward slash to open the insert menu: generated text or images, media files, nested pages, tables, checklists, code blocks and assistant instructions.
- Type /page to nest a sub-page, link an existing page, or generate a sub-page from a prompt.
- Type /visualize followed by a description, and ChatGPT builds a working interactive web app inside the document.
- Type @ to mention teammates, other pages, past chats or the Dot agent.
One /visualize example is a savings calculator with sliders. Starting with $10,000, adding $1,000 a month for 20 years at a 7% return, it projects a balance of $561,314. A planetary orbit simulation tracking Saturn and a small playable piano keyboard were built the same way.
A morning news briefing that builds itself
Add a schedule and a page turns into a dashboard. One example page refreshes every day at 9:00 a.m. Eastern Time. It searches the web for news about ChatGPT, Gemini, Grok and Claude and logs each item with its source, launch date and a priority tag. Below the news sits a social posting queue with ideas for YouTube Shorts, Instagram Reels, TikTok and LinkedIn. For each high-priority topic it drafts a 30- to 60-second video script, platform-specific captions and a list of the visuals needed. A tracking table at the bottom records published URLs, 48-hour view counts, retention, click-through rates and post-mortem notes.
The whole page came from a single prompt in the chat, and ChatGPT finished building it in under four minutes.
▲ A workspace that refreshes every morning
A daily management page for email and calendar
A second page connects to Google Calendar and Gmail. On its first run it read the 100 most recent emails and sorted them into messages that need a reply, follow-up checkboxes, tasks waiting on other people and tomorrow’s agenda. It also flags schedule conflicts, such as a haircut that overlaps a team review.
Building a new page from scratch takes four steps:
- Click New in the upper right, choose Page, and give it a name such as Sponsorship Triage.
- In the prompt bar at the bottom, ask ChatGPT to use the Gmail plugin to find open inquiries, urgent response deadlines and needed follow-ups.
- ChatGPT fills the blank page with tables and checklists: what is due today, replies to send, items waiting on the other side, and expected payments with their invoice status.
- Open the schedule settings and have the page pull fresh email several times a day or on a recurring schedule.
Start with a few simple tables and add detail only as you need it. You can scale the workflow up or down depending on how much control you want.
Zapier MCP: reaching apps without a native plugin
ChatGPT’s built-in plugins come from a small catalog of pre-approved public tools. Niche platforms are often missing; the community platform Skool, for example, has no native plugin. Adding Zapier’s MCP server as a custom server gives ChatGPT access to more than 9,000 apps and over 66,000 triggers and actions, where a trigger is the event that starts a task and an action is the step it performs. The same server also works with Claude, Cursor and custom agents.
Setup follows these steps:
- In Zapier’s MCP dashboard, choose ChatGPT as the target agent and copy the server URL it provides.
- In ChatGPT’s plugin settings, choose Create custom MCP server and enter a name and the URL.
- Review the OAuth sign-in and the risk notice, approve them, and link your Zapier account.
- On Zapier’s server configuration page, attach the apps ChatGPT may use, such as YouTube, Zoom or Google Ads.
Once connected, the agent searches Zapier’s directory for the tool it needs, and when no pre-built action fits, it writes the API call on the fly. The connection itself is free. The setup flow asks you to acknowledge data access risks, and that warning deserves attention: every app you attach widens what the AI can see and change, so connecting only the apps a workflow actually needs appears to be the safer choice.
Platform agents: designing a repeatable worker
The second path runs through the agents section of the OpenAI Developer Platform. It offers templates such as an SRE agent for incident response, an AI teammate for Slack, a data agent, a GitHub issue investigator, and a bulk invoice reviewer that checks invoices against company policy documents. Opening a template exposes the agent’s definition prompt, base model, output format, reasoning effort and verbosity. Turning on subagents lets the main agent hand pieces of a complex job to specialized helper agents.
In the hosted tools section you can enter any MCP server URL or pick from a directory of about 50 built-in servers, including Airtable, Asana, AWS, Box, Cloudflare, GitHub, Google Drive, Jira, Notion and Zapier.
You can also describe the agent you want in plain language. Asked for a daily YouTube analytics tracker that logs views, watch time and subscriber growth to Google Sheets through Zapier MCP, the platform drafted full instructions with scheduled runs, error handling and a JSON output schema. From there you switch on the tools the agent needs, in this case Zapier MCP, Google Drive, Google Sheets, web search and code execution, and start a session in an OpenAI-hosted or self-hosted sandbox, an isolated environment where you can inspect execution traces and debug errors. A Build in Codex option installs the OpenAI Developers plugin so you can develop, test and package the agent’s code in a dedicated coding environment. When you generate an agent this way, spelling out the output format, such as the JSON schema or spreadsheet layout, brings results closer to what you want.
Dot: work that finishes while you are away
The third path is Dot, a personal agent with its own virtual computer. It keeps sessions running in the background and can reach you through chat, phone calls or Slack. The key difference from a regular chatbot appears to be persistence: closing the window does not stop the work. Dot competes with other personal agents such as Muse and Grokbot.
In one test, Dot was asked to use Zapier MCP to pull the last 30 days of YouTube analytics and build a Google Sheets report. Partway through, it noticed that two YouTube channels were connected to the Zapier account and stopped to ask which one to use. After getting an answer, it finished the job while the user was in the kitchen and sent a push notification to the user’s phone. Along the way it fixed a formula reference error in the sheet on its own. The finished workbook had summary metric cards, a daily performance table, a top-videos ranking and a daily views chart, and it added percentage changes and a traffic-source breakdown that the prompt never asked for.
▲ A report finished while the user was away
The run went smoothly partly because of preparation. Before sending the request, the YouTube and Google Sheets connections were checked in Zapier’s Apps tab to make sure both were authenticated. Background agents ask for confirmation through phone notifications, so a missed alert can leave a task waiting.
Choosing where to start
The three paths complement rather than replace one another. Use Pages to collect information you check every day on a single sheet. Use a platform agent when you need the same output format, run after run. Use Dot when a task spans several apps and you want it done while you step away. When a needed app is missing, Zapier MCP widens the reach of all three.
A sensible way to begin:
- Pick one low-stakes task, such as email triage or a news roundup, build it as a page, and watch it for a few days.
- Connect only the outside apps that task needs, and read the OAuth risk notice before approving.
- Confirm that each connected app is authenticated before a run, so jobs do not stall midway.
- Even when an agent fixes small errors by itself, verify its key numbers and outputs before you put it on a recurring schedule.