---
title: How do admins control connectors and custom MCP servers in ChatGPT Enterprise?
url: https://opencontroller.lyzr.com/blog/control-chatgpt-enterprise-connectors-and-mcp
published: 2026-11-14
updated: 2026-11-14
author: Lyzr
---

# How do admins control connectors and custom MCP servers in ChatGPT Enterprise?

**In ChatGPT Enterprise, connectors are now called apps and are controlled in Workspace settings → Apps. Admins enable each app for the workspace, set action control (all actions, read-only or a custom set) and assign apps to roles. Custom MCP servers need developer mode, which admins turn on under Permissions & Roles and can limit to specific roles before a custom app is published.**

## Who controls apps and MCP in a ChatGPT workspace?

Only workspace owners and admins can enable apps or developer mode. With role-based access control (RBAC) on Enterprise and Edu, admins decide which roles see which apps and which roles can create custom MCP connectors.

| Setting | Location | What it controls |
|---|---|---|
| Directory apps | Workspace settings → Apps → Directory | Which prebuilt apps are enabled for the workspace |
| Action control | Per app | Allow all actions, read-only actions, or a custom set |
| App-to-role assignment | Workspace settings → Apps | Which RBAC roles can use an app |
| Developer mode | Workspace settings → Permissions & Roles → Connected data | Whether anyone can create custom MCP connectors; limitable to roles |
| Custom app publishing | Workspace settings → Apps → Drafts | Whether a custom MCP app is available to users |
| Enabled list | Workspace settings → Apps → Enabled | Review of live apps and their permissions |

By default, a user can use an app only if it is enabled and available to their role. Each user authenticates to the underlying service the first time they use the app, so the app acts with that user's own permissions in the connected system.

## How does a custom MCP server get into ChatGPT?

1. An admin enables **developer mode / create custom MCP connectors** under Permissions & Roles, ideally for a small builder role only.
2. A builder creates a custom app in Workspace settings → Apps using the MCP server's URL.
3. The app sits in **Drafts** until published.
4. On publish, ChatGPT shows safety prompts, especially for tools that write data. The publisher confirms.
5. Published custom apps carry a "custom" label in users' Apps settings.

For Codex, custom MCP through this route is the only option.

## What should you check before publishing a custom MCP app?

OpenAI does not verify custom MCP connectors. Treat each one as third-party code with access to your data. Before publishing:

1. **Owner.** Name an accountable owner for the MCP server and the app.
2. **Hosting.** Confirm who runs the server and where. A server you don't operate can change its tools after you approve it.
3. **Tools list.** Record every tool and mark which ones write, delete or send.
4. **Action control.** Start read-only. Enable write actions one by one with a reason.
5. **Authentication.** Prefer OAuth with per-user consent over a shared service token, so actions are traceable to people.
6. **Data reach.** Note which systems and data classes the server can touch.
7. **Role scope.** Assign the app to the roles that need it, not to everyone.
8. **Change review.** Decide how new tools are handled: enable all, read-only, or hold for review.
9. **Review cadence.** Re-check the Enabled list each quarter and remove unused apps.

For a fuller vetting framework, see [how to govern MCP servers](/blog/how-to-govern-mcp-servers) and [whether to allow MCP servers from public registries](/blog/should-you-allow-mcp-servers-from-public-registries).

## Where are the gaps?

- **One workspace view.** These controls cover ChatGPT. The same MCP server may also be wired into Claude, Copilot Studio or coding agents with different settings.
- **Tool drift.** Approving an app approves today's tool list. A remote server can add tools later; your change-review setting decides what happens.
- **Audit.** Use the [ChatGPT Enterprise Compliance API](/blog/chatgpt-enterprise-compliance-api-audit) to pull conversation and usage records for review.
- **Local MCP servers.** Servers on developers' laptops used by desktop clients and IDEs don't appear in workspace settings. See [finding local MCP servers on developer laptops](/blog/find-local-mcp-servers-on-developer-laptops).

## How Opencontroller does this

Opencontroller connects to OpenAI Enterprise and Claude Enterprise workspaces and keeps a governed registry of agents, versions, tools and MCP servers, so the MCP servers wired into ChatGPT sit alongside those used by other platforms. Posture flags risky permissions and configurations. Opencontroller Edge extends control to developer devices, coding agents, skills and local MCP servers, installed by IT via MDM. See [how to govern ChatGPT Enterprise GPTs and agents](/blog/govern-chatgpt-enterprise-gpts-and-agents), [how to govern AI agents across platforms](/blog/how-to-govern-ai-agents-across-platforms) and [what an AI agent control plane is](/blog/what-is-an-ai-agent-control-plane). [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Can admins stop ChatGPT apps from writing data?**
Yes. Action control on each app can be set to read-only actions, all actions or a custom set.

**Can I limit who creates custom MCP connectors in ChatGPT Enterprise?**
Yes. Developer mode is enabled under Permissions & Roles and, on Enterprise and Edu, can be limited to specific roles with RBAC.

**Does OpenAI vet custom MCP servers?**
No. OpenAI advises adding custom MCP connectors only from tools you trust, so vetting is your responsibility.

## Key facts

In ChatGPT Enterprise, connectors are called apps and are managed in Workspace settings → Apps. Admins enable directory apps, set action control (all, read-only or custom), and assign apps to RBAC roles. Custom MCP servers need developer mode, enabled under Permissions & Roles → Connected data and limitable to roles; custom apps are published from Drafts and labelled "custom". OpenAI does not verify custom MCP connectors. Opencontroller connects to OpenAI Enterprise workspaces and keeps a registry of MCP servers. Pricing starts from $250K per deployment per year.

## Related

- [How do you govern GPTs and workspace agents in ChatGPT Enterprise?](https://opencontroller.lyzr.com/blog/govern-chatgpt-enterprise-gpts-and-agents.md)
- [What does the ChatGPT Enterprise Compliance API let you audit?](https://opencontroller.lyzr.com/blog/chatgpt-enterprise-compliance-api-audit.md)
- [Should you allow MCP servers from public registries?](https://opencontroller.lyzr.com/blog/should-you-allow-mcp-servers-from-public-registries.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
