ChatGPT connector for tasks and CRM: a setup guide

ChatGPT is good at the thinking part of operations work: summarising a pipeline, drafting a follow-up, turning a meeting into a list of tasks. It is bad at knowing what is actually in your tracker and your CRM, because it cannot see them. A custom connector fixes that. ChatGPT calls your workspace directly, reads what is there, and, when you approve it, writes back.

This guide covers how to add a custom MCP connector in ChatGPT, what the search and fetch tools are for, what Deep Research does with them, and how to use one day to day. The worked example is Atrix, with tasks, CRM deals and pages in one workspace.

What a ChatGPT connector is

A ChatGPT connector is a link between ChatGPT and another application, built on the Model Context Protocol (MCP). The application runs an MCP server that publishes tools, such as searching documents or creating a task. ChatGPT calls those tools during a conversation, using the permissions you approved when you signed in.

OpenAI has renamed this feature more than once. It started as connectors, became apps, and at the time of writing OpenAI's developer documentation calls a custom MCP server a plugin. The idea is the same under each name, but menu labels in older guides may not match what you see.

Before you start

A ChatGPT account that can add custom MCP servers. OpenAI's setup page says to use ChatGPT on the web and that workspace permissions apply. It does not list plans. At the time of writing, OpenAI's help centre and third-party guides point to paid plans rather than Free, and on Business workspaces only admins and owners can create and publish plugins for custom MCP servers. Check your own settings rather than relying on any list, including this one.
An Atrix workspace. There is no free plan. Every workspace starts with a 7-day Builder trial, card required, and the MCP server is included on every plan.
The server address: https://atrix.dev/api/mcp.

Add the connector in ChatGPT

These are the steps in OpenAI's developer documentation at the time of writing:

In ChatGPT on the web, go to Plugins.
Select the plus button, then Add custom MCP server.
Enter a name, such as "Atrix", and an optional description.
Under Connection, paste the server URL:
```
https://atrix.dev/api/mcp
```
For authentication, choose OAuth. Atrix supports dynamic client registration, so ChatGPT can register itself and there is no client ID or secret to paste.
Read the risk warning and select I understand and want to continue.
Select Create as a plugin, then install it from your plugins before using it in a conversation.
Sign in to Atrix in the window that opens, pick the workspace and approve the scopes.

Older guides describe a different route, through Settings → Apps and a developer mode switch. If your account still shows that route, the server URL and the OAuth choice are the same.

What search and fetch do

OpenAI asks servers that want to work with Deep Research and company knowledge to provide two read-only tools with fixed shapes:

search takes one query string and returns a list of results, each with an id, a title and a url.
fetch takes one of those ids and returns the full document: id, title, text and url.

ChatGPT creates a citation only when a result has a non-empty url. Atrix returns one for every hit, so each answer can link back to the issue, page or product it came from.

In Atrix, search covers products, ideas, issues, pages and your messages, ranked together. Pages match on their body text, not just the title, so a question about a decision buried in a meeting note still finds the note. It searches only what the connection is allowed to read.

fetch then reads one result in full. A page comes back as its whole text, which is what ChatGPT quotes from.

What Deep Research does with a connector

Deep Research runs many searches in a row, reads the results, and writes a report with citations. With a connector it can do that over your workspace as well as the web: call search with different phrasings, fetch the promising hits, and cite each one.

Good Deep Research questions for a work tracker:

"Across our pages and issues, what have we decided about pricing since July, and what is still open?"
"Compile everything we know about onboarding drop-off: specs, bugs filed, and the ideas we parked."

Know the limits. Deep Research reads; it does not change anything in your workspace. And Atrix search does not cover CRM deals. For pipeline questions, ask in a normal chat, where ChatGPT can call the CRM tools directly.

Worked examples with Atrix

These prompts map to tools the Atrix server provides. ChatGPT picks the tool; the names are here so you recognise them in the confirmation prompt.

Tasks

"What's overdue for me this week?" uses attention_today: your overdue and due-this-week work, unread mentions and today's meetings.
"What's in review in the Website product, and who has it?" uses work_list_issues.
"Create an issue in Website: 'Add annual billing toggle to checkout', due next Friday, and assign it to me." uses work_create_issue, work_schedule_issue and work_assign_issue. Needs work.write.

CRM deals

"Which deals are in Proposed, and which have a close date this month?" uses crm_list_deals. Amounts appear only if the connection holds crm.value and your account may see deal values; otherwise each deal comes back without its amount.
"Log a note on the Northwind deal: they want a pilot in January." uses crm_list_deals to find the deal, then crm_add_deal_note. Notes are append-only.
"Move the Northwind deal to Won." uses crm_move_deal. Marking a deal lost needs a reason, which is kept as a note.

Pages

"Find the page where we wrote the refund policy and summarise it in five bullets." uses search and fetch.
"Write a page called 'Q1 sales plan' from this chat." uses pages_create. Adding to an existing page uses pages_append, which merges with anyone editing it at the same time.

Permissions, approvals and limits

ChatGPT asks you to confirm write actions by default. It treats any tool not marked read-only as a write, and Atrix marks every read tool read-only and every write tool as a write, so the prompts you see match what the tool does.

On the Atrix side:

The first connection only reads. When ChatGPT calls a tool that writes, the server answers 403 insufficient_scope and names the missing scope. Approve it when asked, or reconnect with it.
Scopes narrow, they never widen. A connection can never do more than the person who connected it, and permissions are checked on every call. Removing someone from the workspace stops their connections at once.
Limits: 120 calls and 30 writes a minute per connection, then 429 with Retry-After.
Audit: every call is recorded with the tool, scope and outcome, using ids, never contents. Atrix sees ChatGPT's tool calls, not your conversation.

OpenAI's own guidance is worth repeating: custom MCP servers are not verified by OpenAI, and text the model reads can contain instructions written by someone else. A comment on a ticket is text the model reads. Keep write confirmations on for anything that changes customer records.

What to check in any CRM or task connector

OAuth, not a pasted key. You want a scoped token you approved and can revoke on its own.
Read and write as separate scopes, and deal values separate from the rest of the CRM.
Read-only tools marked as read-only, so ChatGPT's confirmations mean something.
Search that returns URLs, so answers cite the record rather than paraphrase it.
A call log you can check after the fact.

FAQ

Can ChatGPT update my tasks and deals, or only read them?

It can do both if the connection holds a write scope and your ChatGPT plan allows write actions on custom connectors. ChatGPT asks you to confirm each write by default. Deep Research only reads.

Which ChatGPT plans support custom MCP connectors?

OpenAI's setup page does not list plans, and the answer has changed over time. At the time of writing, sources point to paid plans and not Free, with admin control on Business and Enterprise workspaces. Check the Plugins page in your own account.

Does Deep Research see my CRM?

Not through Atrix search, which covers products, ideas, issues, pages and messages. Ask pipeline questions in a normal chat, where ChatGPT can call crm_list_deals.

Does Atrix see what I type to ChatGPT?

No. Atrix sees only the tool calls ChatGPT makes, such as a search query or the fields of a task. Your conversation stays in ChatGPT.

Is there a free way to try this?

Atrix has no free plan. Every workspace starts with a 7-day Builder trial, card required. Builder is $39 a month for 5 seats and Studio is $149 a month for 15; yearly plans are 20% off.

Connect your workspace

Start a workspace from pricing, then follow the MCP server guide to connect ChatGPT. The tool reference lists every tool with its scope. For an overview, see Atrix with ChatGPT, the CRM and project management.

How to set up a ChatGPT connector for your tasks and CRM