Connecting ChatGPT and Grok: What Worked, What Failed, and What Comes Next

I have been working with two AI assistants: Bob on Grok and Dot on ChatGPT. Each can help me think, write, investigate, and review work. But getting them to work together introduced a surprisingly ordinary problem.

I was the messenger.

I would take a response from one, copy it into the other, carry the critique back, and repeat. That works for an occasional second opinion. It becomes cumbersome when you want an ongoing conversation around a project.

I wanted a shared place where Bob and Dot could leave notes for each other, with me still deciding what happens next.

We now have a working shared notebook. Getting there taught me as much as the successful connection did.

What we built

The notebook is a small service hosted on Cloudflare. It uses MCP, the Model Context Protocol, to give the assistants tools for reading and adding entries.

Think of it as a shared project notebook with a record of who wrote each note and when. Bob can leave an observation. Dot can read it and add a response. The conversation stays together instead of being scattered across copies in separate chats.

The notebook uses an authenticated connection. It does not require me to publish my conversations or give one assistant unrestricted access to the other.

Our first test was deliberately simple. Bob posted a harmless connection-check message. Dot read it, replied in the same thread, and then read the thread again to verify that her reply had saved.

That last step mattered. Asking an assistant to write something and seeing the saved result are two different tests. We checked the result.

Where we stumbled

The first difficulty was getting the right kind of connector into the right interface.

An early setup led to a plugin page that pointed toward a desktop application instead of giving us the web sign-in flow we needed. We had to correct the connection setup for ChatGPT on the web.

Then the sign-in process failed in a way that looked like a typing mistake.

I thought I had entered the passphrase incorrectly. The page closed, and trying again led to an expired request. Bob later reported that the passphrase had worked. A security setting on the approval page had prevented the browser from returning to ChatGPT. Clicking Approve again tried to reuse a request that had already been consumed.

Bob fixed the approval page and reported testing the complete sign-in flow in a real browser. A fresh connection attempt then allowed us to continue.

The lesson for me was practical: test the complete user journey. A server can respond correctly while the person using it still cannot finish signing in.

The options we considered

We could have continued copying and pasting. That remained the simplest fallback, but it left me responsible for moving every message.

We also discussed ways to host and expose the notebook, including local testing and a hosted service. We chose a hosted notebook so the assistants could reach the same record without depending on a temporary local session.

Authentication was another decision. For our private connection, we used OAuth sign-in rather than treating the notebook as an open endpoint.

Finally, we had to separate two questions:

Can the assistants exchange messages?

Will they notice new messages without me prompting them?

The notebook test showed that Dot could read Bob’s note and save a reply. For the second question, we considered scheduled checks and event-driven notifications. I asked Dot to check hourly, and that watch has been configured.

At this stage, the message exchange has been tested and the hourly monitoring configured. I still want to see the monitoring work through its normal cycle before calling the whole process finished. We also haven’t yet confirmed that Bob has read Dot’s reply.

We have not established an automatic, continuous conversation in which both assistants wake each other and reply indefinitely.

Why this could be useful

The benefit I want is better work with less effort spent carrying information between tools.

One assistant could prepare a draft while the other reviews it against the original brief. A review could focus on missing assumptions, unclear language, weak evidence, or questions a customer might ask.

For a business process, one could outline the workflow and the other could look for exceptions: What happens when a required field is missing? Who handles a disagreement? Where does a person need to make a decision?

For a small experiment, the notebook could hold the proposed test, the concerns raised before running it, and the results afterward. That would make it easier to follow how a conclusion was reached.

For content, Bob and Dot could offer different perspectives before a post comes back to me for approval.

These are possible uses, not results we have already proved. Our completed test demonstrates that messages can be read and saved through the notebook. The next step is to find out whether that exchange improves real work.

Two assistants still need judgment

A second AI opinion is useful, but agreement is not proof.

Both assistants can miss the same fact or accept the same faulty assumption. Important claims still need evidence. Decisions still need someone accountable for making them.

There is also a difference between a collaborative review and an independent test. If one assistant can see the other’s answer, that is a useful discussion, but it is no longer a blind comparison. We need to choose the right setup for the question.

I also want clear boundaries. A notebook message should not, by itself, authorize a purchase, a public post, or a change to a business system. Sharing a workspace does not mean sharing every private conversation.

For this experiment, the initial messages contained no confidential information. This article leaves out credentials and private connection details as well.

What success means so far

We got past the connection problems. Dot read Bob’s message, wrote a reply, and verified that it was saved. We also configured an hourly check to reduce the need for me to keep asking whether a new message has arrived.

That is a useful foundation.

My broader interest is building an intelligent company: using AI to support everyday work while keeping human judgment, evidence, and responsibility in the process.

This small project gives us a way to try that with two assistants working from a shared record. Now the worthwhile question is what they can help us do better, and how we will test it.

Leave a Reply