ChatGPT connectors guide
ChatGPT connectors for enterprise: admin and security guide
How ChatGPT connectors and custom MCP connectors work for business plans: workspace controls, write-action confirmation, and connecting ERP, CRM and HR data.
Short answer
ChatGPT connectors let ChatGPT search and act on data in other applications. Custom connectors are remote MCP servers added by URL, with OAuth sign-in. In business workspaces, workspace permissions and security settings decide who can add and use them, and ChatGPT asks for confirmation before write actions. For enterprise data, pick connectors where users sign in with their existing access, so ChatGPT only sees what each user is already authorized to see.
Updated
How ChatGPT connectors work
A connector gives ChatGPT tools from another service: search files, fetch a record, run a query. OpenAI has used the names connectors, apps and plugins for these over time, so check OpenAI’s current help pages for the exact menu labels in your workspace.
Custom connectors are remote MCP servers. OpenAI’s developer documentation describes adding one from ChatGPT on the web with a name, the server URL and an authentication method: OAuth, or no authentication for public data. It supports SSE and streamable HTTP transports, and OAuth clients can be registered with static credentials, Client ID Metadata Documents or dynamic client registration (OpenAI: add a custom MCP server).
OpenAI also notes that a server supporting ChatGPT deep research and company knowledge implements read-only search and fetch tools, while other custom servers can expose any tools they need (OpenAI: building MCP servers).
What ChatGPT workspace admins control
OpenAI’s developer guide states that workspace permissions and security restrictions apply to adding and using custom MCP servers (checked 2026-11-19). In practice, an enterprise rollout decides four things:
- Who can add custom connectors. On business plans this is usually limited to admins or a named group, so the workspace does not collect unreviewed MCP servers.
- Which connectors are approved. Keep a short list per department, each with an owner, and remove connectors nobody uses.
- How writes are handled. OpenAI documents that write actions require confirmation by default, and that tools are treated as writes unless the server marks them read-only. Decide which connectors may write at all.
- Which data the server reaches. OpenAI advises connecting only to MCP servers you trust, because an untrusted server can access information shared with it. Review each server’s operator, authentication and data handling like any other vendor.
Connecting ChatGPT to ERP, CRM, HR and data platforms
File and mail connectors answer “where is that document”. Most business questions need systems of record: bookings from the CRM, revenue from the ERP, headcount from the HR system, usage from the warehouse. Three problems show up as soon as ChatGPT reaches them.
- Permissions. A connector built on one integration credential returns what that credential can read, to everyone. Finance, HR and customer data need user-level access: role-based access, row- and column-level security and sharing rules from each source system.
- Definitions. “Revenue” in NetSuite, Salesforce and the warehouse are three different numbers. Without approved definitions, any assistant may pick one without saying which.
- Evidence. A number in a chat is only useful if someone can check it. The answer should show the source, the query and the definition used.
This is the gap a governed connector fills. Typical first systems are Salesforce and HubSpot for revenue teams, NetSuite, SAP S/4HANA and QuickBooks for finance, Workday for HR, ServiceNow for IT, and Snowflake or Databricks for the data team. The connector catalog lists every system by category.
Checklist before you approve a ChatGPT connector
- Confirm how users authenticate: their existing account in the source system, or the company’s SSO, never a key pasted into ChatGPT.
- Test with two users who have different roles in the source system and compare answers. They should differ exactly where their permissions differ.
- List the connector’s tools and mark which can write. Keep writes off until a team has a workflow and an owner for them.
- Ask the same question in ChatGPT and in the source system’s own report. Where they differ, find out which definition each used.
- Check where requests and answers are logged, and who in security can review them.
Using Nexatron as a ChatGPT connector
Nexatron’s hosted MCP server works as a ChatGPT custom connector. Where your workspace allows custom connectors, add https://api.nexatron.io/api/v1/mcp, choose OAuth, and sign in. Nexatron supports OAuth 2.1 with PKCE and dynamic client registration, so no client ID or secret is needed.
Users sign in with their existing account in each system, or through the company’s SSO if it has one, and Nexatron inherits their existing permissions, so every answer contains only data they’re already authorized to see. Questions use your approved definitions, each answer shows the definition, source and query behind it, and governed answers are written to an audit trail. The same connection works in Claude, Microsoft 365 Copilot and Gemini, so every assistant gives the same number. Step-by-step setup is on the ChatGPT setup page.
Frequently asked questions
- What are ChatGPT connectors?
- Connectors let ChatGPT use tools and data from other applications. Custom connectors are remote MCP servers added by URL, so any system with an MCP server, or a governed layer such as Nexatron, can be reached from ChatGPT.
- How do I add a custom connector to ChatGPT?
- In ChatGPT on the web, open the settings for connectors or apps, choose to add a custom MCP server, enter its URL and choose OAuth. In business workspaces an admin may need to allow it first. OpenAI’s current help pages have the exact labels.
- Can ChatGPT connectors change data in our systems?
- Connectors can expose write tools. OpenAI documents that ChatGPT asks for confirmation before write actions by default. Nexatron goes further: questions never change data, and actions stay off until a workspace admin enables them for a connection.
- Do ChatGPT connectors respect our existing permissions?
- That depends on the connector. If each user signs in with their existing access, the source system’s permissions apply. If the connector uses one integration credential for everyone, they do not. Nexatron inherits each user’s existing permissions in every connected system.
- Can ChatGPT connect to a database or data warehouse?
- Yes, through an MCP server in front of it. For enterprise use, the server should query with the user’s access so row access policies and masking apply, and should return the SQL it ran. Nexatron does this for Snowflake, Databricks, BigQuery, Postgres and other platforms.
- Is there a difference between ChatGPT connectors and Claude connectors?
- Both use MCP for remote connectors, so one MCP server can serve both. Admin controls, menu labels and plan availability differ by vendor. Nexatron uses one endpoint for Claude, ChatGPT, Copilot and Gemini.
Sources
Statements about other products come from these public pages, read on November 19, 2026.
Put governed business data in ChatGPT
Tell us which systems your teams ask about. We show governed answers in ChatGPT under each user’s existing permissions.