ChatGPT Connectors API: How the Different Connection Routes Work
ChatGPT offers three distinct routes for connecting to outside tools and data: interface connectors, MCP connections, and direct API calls. Here is how they differ, which to use for which job, and how fenbs illustrates shared project context through a supported connection.
Updated 5 min read
ChatGPT supports three distinct ways to connect external tools and data: connectors available through the interface, MCP connections that follow an open standard, and direct API calls your application makes. These are not interchangeable names for the same thing. Choosing the right route for your use case saves considerable debugging time. Tools like fenbs connect through MCP, giving ChatGPT access to shared project context through a documented, reviewable connection rather than a guessed endpoint.
Understanding these routes also matters for decisions: which tools you need to evaluate, which require admin approval, and which require a developer to set up.
Interface connectors: approved integrations in the ChatGPT UI
Interface connectors are the integrations you add through the ChatGPT settings or during a conversation: file storage services, productivity tools, and specific apps that OpenAI has added to the interface. You sign in or authorize the connection through a browser, and ChatGPT can then read or act on data from that service during a conversation.
These are curated integrations. OpenAI decides which services appear in the connector list. If the tool you need is not on the list, this route is not available to you. See ChatGPT connectors and apps for what is currently available. Troubleshooting steps are in ChatGPT connectors not working.
MCP connections: the open standard route
MCP, or Model Context Protocol, is an open standard that lets an AI app call any tool that implements the standard. ChatGPT supports MCP connections through developer settings. This means any organization can publish an MCP server, and ChatGPT can connect to it without OpenAI adding it to a curated list.
In ChatGPT, MCP connections are added through developer or advanced settings; which plans support this is OpenAI's to say, so check their current documentation (opens in a new tab). The connection uses a browser sign-in for authentication rather than a pasted token, and the server declares its available tools in a standard format the AI can read.
fenbs connects this way. When you authorize the fenbs connection, ChatGPT can read your projects, tasks, Decisions and rules and Where we left off as standard tools. The ChatGPT connectors for Jira article covers a similar MCP-based integration as a point of comparison.
Shared project context through a documented MCP connection
fenbs publishes its tools through MCP, so ChatGPT can read the current project context without a custom integration or a guessed endpoint. Connect it from Integrations and the assistant reads your projects, tasks and Decisions and rules through the standard tool interface. Every read and write is attributed in History.
Create a free accountFree for one person · No credit card · Works with Claude, ChatGPT, Cursor and more
Direct API calls: your application talks to ChatGPT
The ChatGPT API, meaning the OpenAI Chat Completions or Responses API, is not a connector; it is the other direction. Your application sends messages to ChatGPT and receives responses. You are the caller, not the receiver.
This is the route for building products that use ChatGPT as a backend: a customer support tool, a document processor, a coding assistant you build yourself. The AI does not "connect" to your tool; your tool calls the API. See the API documentation for the available endpoints.
When decisions live outside the connection
A connection gives an AI access to data. It does not give the AI the decisions people made about what to do with that data. Which customers should be contacted first, which code patterns are approved, which topics the AI should avoid: those decisions need their own home.
Building a record of those decisions in fenbs means any connected AI reads them at the start of a session, whether it is accessing a Jira board, a SharePoint folder or a CRM. The context travels with the work, not with a specific connection.
How to do this with fenbs
- Create a free account and add a project with the decisions and rules your AI needs to follow.
- Choose the connection route that fits: interface connector for a curated tool, MCP for any MCP-compatible service including fenbs, or direct API if you are building a product.
- For MCP, go to Integrations and follow the ChatGPT connection guide. Authorize the fenbs connection with your account.
- In Decisions and rules, record the instructions that apply to this project: what the AI may change, what it must ask about, and what is out of scope.
- Before each session ends, ask the AI to save progress notes and update Where we left off.
Choosing the right route for your use case
Use interface connectors when you need a curated, well-supported integration with minimal setup. Use MCP when you need flexibility, want to connect a tool not on the curated list, or want the same server to work across multiple AI clients including Claude and Cursor. Use the direct API when you are building a product, a workflow or an automation where your own application needs to control the prompt and the output.
For teams, the most durable setup usually combines all three: interface connectors for the productivity tools people already use, an MCP connection for shared project context through a tool like fenbs, and direct API calls for any automated steps that run without a person in the loop.
When these routes are not enough
If the tool you need is not available as an interface connector and does not have an MCP server, you have two options: build an MCP server for it yourself (see the MCP server boilerplate as a starting point), or accept that ChatGPT cannot reach it directly. fenbs does not bridge arbitrary internal systems; it holds the work context you add to it.