MCP for ChatGPT Module Included
Like Claude, ChatGPT registers itself — no pre-registered client, no secret. A user adds the connector and signs in to Vaks in the browser. Two differences from Claude: ChatGPT's connector UI is still evolving, and it does not surface the pre-built workflow menu, so you recreate those yourself.
ChatGPT's connector interface is in beta and moves. The exact place to add a custom MCP connector (developer mode / connectors) and its labels changed through late 2025. Use OpenAI's current documentation for the precise steps in your plan; the Vaks-side values below do not change.
Before you start
Everything ChatGPT-specific is on this page. The shared parts live in the hub:
- Enable the connector for your organization — Enable and scope it. Off by default; consent is refused until it is on.
- Permissions model — an assistant's rights are the user's rights. Read-only until you turn writes on.
- The URL to hand out — the tenant's MCP URL, or the public MCP URL if a gateway was deployed (reaching it from outside).
Connect ChatGPT
ChatGPT self-registers via Dynamic Client Registration — no pre-registered client, no secret. In outline:
- In ChatGPT, add Vaks PM as a custom connector (via connectors / developer mode — see OpenAI's docs for the current location).
- Paste the MCP server URL (
https://<tenant>/mcp, or your public MCP URL). - ChatGPT discovers how to authenticate and opens a browser sign-in to Vaks (local or SSO).
- The user approves the consent screen and ChatGPT starts working — standard OAuth with PKCE, one-hour tokens that refresh automatically.
Per-user by nature. Each ChatGPT user signs in as themselves, so the assistant acts as that person, bounded by their own permissions. In Vaks, Admin → Integrations → MCP → Active AI connections shows one connection per user.
Pre-built workflows
Vaks exposes a set of pre-built workflows as MCP prompts — see the full library in the hub. Claude shows them as a menu automatically; ChatGPT, like Copilot Studio, does not surface that MCP primitive today. Because a prompt is only a pre-written instruction, you recreate it as text:
- A prompt library your users copy-paste — one line per workflow (e.g. “Review all my projects and give me a prioritised list of risks”, “Write a shareable status report for project <name>”). Reuse the example texts from the Copilot Studio guide.
- A custom GPT whose instructions and starter prompts embed these workflows, so they appear as suggestions.
Write workflows need writes enabled. The workflows that end in a write only function if Allow AI write actions is on. ⚠️ ChatGPT's handling of MCP prompts may change — confirm current behaviour on OpenAI's side.
Troubleshooting
| Symptom | Cause & fix |
|---|---|
| ChatGPT cannot reach the server | Almost always the URL: on a gateway deployment the admin-panel URL is not the public one. Confirm the public hostname — see reaching it from outside. ChatGPT is a cloud service, so the endpoint must be reachable over the internet. |
| The connector option isn't where the docs say | ChatGPT's connector UI is in beta and has moved. Check OpenAI's current documentation for developer mode / connectors. |
| Consent screen refuses, no approve button | The connector is disabled. Enable it in the admin section. |
| Only read tools available | Writes are off, or the connection predates turning them on. Turn on Allow AI write actions and reconnect. |
Related: MCP overview (hub) — permissions, admin toggles, deployment · Copilot Studio · Claude · all integrations.