ChatGPT apps vs Claude connectors: which assistant can actually work in your tools

The short version: same protocol, different gates

Both assistants speak the Model Context Protocol. A work tool runs an MCP server; the assistant discovers its tools ("search", "create issue", "move deal"), signs in with OAuth, and calls them during a conversation. A server built once can serve both, which is why most work tools now list ChatGPT and Claude on the same integrations page.

Where they differ is everything around the protocol: what each vendor calls the feature, which plans can add a server that is not in the directory, who in a company may do it, and what happens before the assistant writes.

What changed in names: connectors, apps, plugins

OpenAI renamed ChatGPT connectors to apps in December 2025, to cover both interactive apps and the integrations that search your data. Since then its interface has moved again: as of October 2026 OpenAI's help centre sends you to Settings → Plugins, and describes a plugin as a package that can hold one or more apps, skills, or both. An app is still the thing that connects ChatGPT to an outside service. Older guides that say "Settings → Connectors" or "Settings → Apps" describe earlier menus.

Anthropic still says connectors. Claude's directory page is titled "Connectors and plugins"; it listed 1,023 connectors when we checked in October 2026, filterable as Web, Desktop extension or Interactive, with plugins on a separate tab.

So when this post says "ChatGPT app" and "Claude connector", read both as "an MCP server the assistant is allowed to call".

ChatGPT apps vs Claude connectors compared (October 2026)

Sources are listed at the end. OpenAI's plan rules come from its article on connecting custom MCP servers; Claude's from its custom connectors article and pricing page.

Plan rules: the gate that decides most teams

Claude is the more permissive of the two. Anthropic's help centre says remote MCP custom connectors work for users on Free, Pro, Max, Team and Enterprise, with Free limited to one custom connector. That means a two-person team on Pro can connect its tracker, CRM and calendar today, with write tools, without asking anyone.

ChatGPT gates by plan and role. OpenAI's article says custom MCP server connections with full MCP support, write and modify actions included, are for Business and Enterprise/Edu customers on ChatGPT web. Pro users can connect custom servers with read and fetch permissions only, and can build apps with the Apps SDK. The article does not list Free, Go or Plus for custom servers. Apps that OpenAI has approved for its directory are a separate matter: their availability depends on the app, your plan, region and workspace.

In practice: if your team is on ChatGPT Plus or Pro and you want ChatGPT to change things in a tool that is not in OpenAI's directory, you need a Business workspace. If you are on Claude Pro, you do not.

Admin control and who can connect what

Both vendors put company admins in charge on business plans, in different ways.

On ChatGPT Business, only admins and owners can create and publish a custom MCP server for the workspace. Enterprise and Edu admins can grant that right to selected members by role, choose which groups see each app, and switch individual actions on or off before and after publishing. One detail matters more than it looks: once published, ChatGPT uses a frozen snapshot of the server's tools. If the vendor adds or changes a tool, nothing changes for your team until an admin refreshes it, and a call to a changed tool can fail without notice.

On Claude Team, Owners and Primary Owners add a custom connector for the organisation; each member then connects their own account to it, so the assistant acts as that person. Enterprise adds a custom role for the same job. Owners can also disable tool calls that render interactive connectors.

For a team of five, Claude's model is less ceremony. For a company that wants every write action reviewed before anyone can use it, ChatGPT Enterprise's per-action control is stricter.

What the assistant can write, and how it asks first

Both assistants can write when the server and your plan allow it: create a task, update a deal, post a message.

ChatGPT gives each app a permission level. Always ask, Allow read actions and Allow low-risk actions are the standard choices; Allow all actions exists for eligible apps and lets supported actions run without another confirmation. OpenAI says especially risky actions may be blocked rather than offered for approval, and admins see risk warnings when enabling apps that can write. Deep research can use custom apps, but only to read.

Claude asks per tool call. You approve the call, or choose "Allow always" for a tool you trust; Anthropic's advice is to keep that for servers you trust to run unsupervised. Tools you do not need can be switched off from the Search and tools menu.

Our rule for either: grant read scopes for the first week, and turn on writes once you have seen what the assistant gets wrong.

Auth and setup differences for tool builders

If you build the MCP server, both assistants want OAuth, with small differences.

• ChatGPT requires an OAuth 2.1 flow that follows the MCP authorization spec, with PKCE using S256. Your server should publish protected resource metadata at /.well-known/oauth-protected-resource. ChatGPT supports dynamic client registration and prefers Client ID Metadata Documents when your server supports them. Ask for refresh tokens (offline_access for OpenID Connect providers), or users will be signed out when the first token expires. The search and fetch tools are no longer required, but company knowledge only uses apps that have them.
• Claude supports OAuth with automatic client registration, or fixed credentials in request headers for servers that use API keys. The server must be reachable on the public internet from Anthropic's IP ranges.

A server that implements OAuth 2.1 with dynamic client registration and protected resource metadata works in both without a client ID to paste.

Where Atrix fits: one server, both assistants

Atrix is the workspace we build, so read this as one example rather than a recommendation. Its MCP server at https://atrix.dev/api/mcp exposes 85 tools over OAuth 2.1 with dynamic client registration, covering work and cycles, pages, CRM, calendar, hiring, people, goals and time. It implements search and fetch, so ChatGPT can cite workspace pages in answers and deep research. It is listed in the Official MCP Registry and on Glama.

Honest status as of October 2026: in Claude you add Atrix as a custom connector by URL; its Claude directory listing is under review, not live. In ChatGPT you add it as a custom MCP server, which means write tools need a Business, Enterprise or Edu workspace, and Pro gets read-only. Tool calls from either assistant do not use Atrix AI runs. Atrix has no free plan: a 7-day trial with a card, then Builder at $39 a month for 5 seats or Studio at $149 for 15. Setup steps are on the ChatGPT integration and Claude integration pages.

How to choose for your team

1. List the tools you want the assistant to write to. If they are all in OpenAI's directory, ChatGPT's plan gate matters less. If one is a custom server, check the rule above.
2. Check your current plan. Claude Pro or Team covers custom connectors with writes. ChatGPT needs Business or above for the same.
3. Decide who approves connections. Want an admin to vet every tool and action? ChatGPT Enterprise does that best. Want each person to connect their own account with little ceremony? Claude does.
4. Test with read scopes first. Ask for a list before asking for a change, in either assistant.
5. Plan for tool updates. On ChatGPT Business and Enterprise, put "refresh the app" on someone's list when a vendor ships new tools.

If you are wiring up Claude, our guide to Claude connectors for small teams picks ten by job; for ChatGPT, see setting up a ChatGPT connector for tasks and CRM.

FAQ

Are ChatGPT apps and Claude connectors the same thing?

They do the same job on the same protocol. Both let the assistant call tools on a remote MCP server with your permissions. The differences are naming, plan rules, admin controls and how each asks before a write.

Did OpenAI rename connectors to apps?

Yes, in December 2025. As of October 2026 OpenAI's interface groups apps under Plugins, where a plugin can bundle one or more apps and skills. Menus in older guides may not match what you see.

Can I use a custom MCP server on ChatGPT Plus?

OpenAI's help article on custom MCP servers lists Business and Enterprise/Edu for full support and Pro for read-only. It does not list Plus, so as of October 2026 assume you cannot, and check your own settings.

Do Claude connectors work on the free plan?

Yes. Anthropic's help centre says custom connectors work on Free, Pro, Max, Team and Enterprise, with Free limited to one custom connector.

Which is better for connecting work tools, ChatGPT or Claude?

For a small team that wants write access to tools outside the directories without upgrading, Claude, because custom connectors with writes are on every plan. For a company that wants admins to approve each tool and action, ChatGPT Enterprise's controls are stricter. The protocol is the same, so a good MCP server works in both.

Next step

If you build or choose work tools, start with what MCP is, then see what Atrix's MCP server exposes and the developer docs for MCP.

Sources

Checked in October 2026: OpenAI, connecting custom MCP servers in ChatGPT, OpenAI, connected apps in ChatGPT, OpenAI Apps SDK authentication, Anthropic, custom connectors using remote MCP, Claude pricing, Claude connectors directory.