The ChatGPT Work guide
ChatGPT Work, connected to your customer support
ChatGPT Work is OpenAI’s answer to getting real work done, and connectors are what make it real. Give it SiteGPT and your support operation becomes something you describe: triage, retraining, reports, one workspace.
> What did our support bot miss this week, and what should we teach it? SiteGPT: 214 conversations, 19 unanswered 4 recurring gaps found, answers drafted from docs Queue new sources for training? approve ✓ Training queued. Report ready to share.Agents read this page too
First principles
What is ChatGPT Work?
ChatGPT Work is OpenAI’s push to make ChatGPT the place where work gets finished, not just discussed: multi-step tasks planned and executed, your company’s context in reach, and external tools available through connectors. It is the business-grade side of ChatGPT, and it changes what a support team can delegate.
The pattern will feel familiar if you have watched the agentic wave: Anthropic ships it as Claude Cowork, OpenAI ships it inside ChatGPT. Both stand or fall on the same question: what can the assistant actually reach? A workspace with no connectors is a very articulate spectator.
That is the part this guide is about: wiring ChatGPT Work into a real, running business system, with customer support as the worked example.
Connectors, explained
How ChatGPT connectors work
Connectors plug ChatGPT into products outside itself: your files, your systems, your platforms. Under the hood they run on MCP, the Model Context Protocol, the open standard the whole assistant ecosystem has settled on. That standardization is quietly the best part: a well-built MCP server serves ChatGPT, Claude, Cursor, and every other client from one implementation, so your connected stack is not a bet on one vendor.
What separates a connector you can trust with customer data from a demo: OAuth approval instead of pasted API keys, read operations separated from writes so the assistant must ask before changing anything, and grants bounded by your own role in the connected product. SiteGPT’s server was built to that bar, with three compact tools (search, execute_read, execute_write) whose annotations tell the assistant exactly which is which.
The support use case
Customer support, run from ChatGPT
The division of labor that works: a SiteGPT chatbot, trained on your own content, answers visitors on your website around the clock. ChatGPT, with the connector, is where you run that front line.
The weekly review, in a sentence
"What did the support bot miss this week?" ChatGPT pulls the unanswered conversations, clusters the recurring questions, and drafts answers from your docs.
Retraining without the chore
Approve the drafts and the new sources queue for training. The bot facing your customers gets smarter without a dashboard session.
Escalations, surfaced
Conversations waiting on a human, sorted by wait time, each one summarized before you step in.
Reports where you work
Volume, deflection, captured leads: live numbers in the format your Monday needs.
Two minutes to connect
Connect SiteGPT to ChatGPT
- In ChatGPT’s connector settings, add SiteGPT with the server URL https://sitegpt.ai/mcp.
- ChatGPT opens SiteGPT’s approval page in the browser: sign in, review the permissions and chatbot access, and connect. No API keys anywhere.
- Ask: "List my SiteGPT chatbots, then show this week’s unresolved conversations."
No SiteGPT account yet? An agent can build and train a chatbot from your website URL with no signup: agent-first onboarding hands back a preview link to claim.
Steal these
Five prompts to start with
Everything below runs against your live account the moment the connector is approved. Reads run freely; anything that changes data asks first.
“List my SiteGPT chatbots and flag any with failed training sources.”
A live inventory with the ones needing attention, from one read.
“Show every conversation this week where the support bot could not answer.”
The knowledge-gap list, clustered by topic, with links into the dashboard.
“Draft answers for those gaps from our docs and queue them for training.”
ChatGPT drafts, you approve, the write goes through the same audited API.
“Which conversations are waiting on a human, and for how long?”
The escalation queue, sorted by wait, each thread summarized.
“Summarize support volume, deflection, and captured leads for the leadership update.”
Live numbers in your format, ready to paste.
The other workspace
ChatGPT Work vs Claude Cowork
They are the two major managed agentic work surfaces, and they rhyme: OpenAI’s lives inside ChatGPT, Anthropic’s is the Cowork desktop workspace. Both plan multi-step work, both check in for approvals, and both reach external tools over MCP.
That last part is the practical takeaway. Because both speak the same protocol, tool builders serve both, and choosing a workspace does not lock your connected stack. SiteGPT connects to each from the same account: one click from the directory on the Claude side, the server URL on the ChatGPT side. If you are genuinely torn, connect both and compare them on your real support workload for a week; the better fit tends to announce itself.
One more option deserves naming: both workspaces still need you in the chair. For work that should run on a schedule whether or not anyone is sitting anywhere, MissionControlHQ (from the same team as SiteGPT) employs a squad of persistent agents, and they operate SiteGPT over these same surfaces.
The deeper comparison, including Claude Code and OpenClaw, lives in the Cowork guide.
ChatGPT Work, answered
- What is ChatGPT Work?
- ChatGPT Work is OpenAI’s offering for getting real work done in ChatGPT: the business-grade side of the product, where ChatGPT plans multi-step tasks, works with your company’s context, and reaches external tools through connectors. Where a chat answers questions, ChatGPT Work is aimed at finishing tasks.
- What are ChatGPT connectors?
- Connectors plug ChatGPT into outside products so it can read live data and take actions there. They run on MCP, the Model Context Protocol, the same open standard Claude and most agent runtimes use. One well-built MCP server, like SiteGPT’s, serves every assistant that speaks the protocol.
- How do I connect SiteGPT to ChatGPT?
- Add SiteGPT as a connector using the server URL https://sitegpt.ai/mcp. ChatGPT opens SiteGPT’s browser approval page: sign in, review the permissions and chatbot access, and connect. No API key is pasted anywhere, and you can revoke the grant at any time.
- What can ChatGPT do with SiteGPT connected?
- Run your customer support platform in plain language: list chatbots, pull unresolved conversations, find questions your support bot could not answer, queue docs for retraining, update custom responses, and summarize volume, deflection, and captured leads.
- ChatGPT Work vs Claude Cowork: what is the difference?
- They are the two major managed agentic work surfaces: OpenAI’s lives in ChatGPT, Anthropic’s is the Cowork desktop workspace. Both reach external tools over MCP, which means tool builders like SiteGPT serve both, and teams can pick either without losing their connected stack. If you are evaluating the pair, connect the same SiteGPT account to both and compare on your real support workload.
- Is my customer data safe in this setup?
- The connector is bounded by an OAuth grant you approve: it can never exceed your own SiteGPT role, token scopes, or chatbot restrictions. Read operations are separated from writes, and anything that changes data asks first. Revoking the grant, from either side, ends access immediately.
- Does SiteGPT work with the rest of the OpenAI tooling?
- Yes. Codex connects over stdio through the SiteGPT CLI (npx -y @sitegpt/cli mcp), and the REST API and agent skill cover anything custom. The Codex guide has the config.
Put support on ChatGPT’s desk
One server URL, one browser approval, and your support operation becomes something you describe.