Look, I'll spare you the suspense and just let you know right now: Dots are amazing. If you want to read on to figure out exactly why I've come to this conclusion, then certainly the ride is, I think, interesting, but I didn't want to bury the lede. If you don't already have the $100 ChatGPT plan, then I highly recommend getting it and using Dots.
I put this thing through its paces. I connected it up to over half a dozen of my own custom software solutions that I built over the last year and a half. I even used it to actually create the MCP servers that it eventually connected itself to. I've been able to successfully get it to manage a full-blown development pipeline with my guidance. It is keeping track of an insane number of tasks, tracking the progress in our own kanban board. It is keeping up with my emails, which has been clutch in doing research prior to client meetings, and has otherwise just been a phenomenal working partner.
But probably the most impressive thing that it was able to do was monitor a Slack channel that I shared with a client and help guide that client towards some decisions on how they want to actually run their business, so we could get a development project kicked off and done. It made it to where I could just observe the conversation as it was unfolding and chime in here and there where I felt like I could actually help steer the discussion. Ultimately, my Dot, whose name is Riri, was able to just handle it more or less autonomously.
Now, that's not to say there weren't some rough edges. There certainly were, although I also have to give this thing a ton of credit, given that I started using it as soon as it was available. It launched on Tuesday and was broadly available to ChatGPT Pro users on Wednesday. It was a remarkably smooth ride, all things considered.
What Even Is a "Dot"?
OpenAI released ChatGPT "Dots" as part of its Dev Day festivities on September 29, 2026. Now, say what you will about that Dev Day. If you watched it, then it definitely left a lot to be desired, but Dots and the release of GPT-6.1 Sol were definitely good things to come out of Dev Day. Had they not overhyped it to the extent that they did, this hedging would not be necessary. I did previously write about how I just wasn't all that excited for Dev Day, and I feel like I should have been.
At any rate, A Dot is a personal assistant that is AI-powered and able to connect to many of your common tools and even some of your custom tools. It is able to take actions on your behalf if you want it to, or you can actually give it its own autonomy to take action on its own behalf. Dots differ from a standard chatbot in that they don't necessarily wait for you to make a request of them. They're a system that can wake up and take action based upon goals that you have set for them in the past, so it's more of a goal- or role-oriented type of assistant rather than a task-based agent.
Within the last couple of months, there have been several offerings put out by the larger labs that are starting to make this space a little crowded. For instance, xAI has released Grokbot, and Meta has released Muse. Each of them has its pluses and minuses. I would argue that all of them are better than the open-source alternatives that sort of paved the way for these to exist. If you have been keeping up in the space, then you may have heard of those open-source versions, which came out a few months back, named OpenClaw and Hermes. All of these solutions are meant to tackle roughly the same problem space, although, in my opinion, Dots are probably the most well-rounded of the group.
The Setup Was Painless
The setup was remarkably easy, and I say that because I set up OpenClaw and Hermes, and both of those required a lot of customization in order to get right. With Dots, it was literally just clicking on the icon to set up your Dot, then giving it a name, and then you're off to the races. From there, certainly, you're going to want to connect it up to as many services as you possibly can. The more it has access to, the more useful it can be for you.
One thing that really struck me, though, is the fact that even if you don't have all of the direct connections set up, by and large, because OpenAI models are just so good at browser use and computer use, if it can't get it through an MCP server, then it can be quite effective at getting everything that you need by using the browser. It can do that on its own cloud machine or, if that's blocked, it can actually use Chrome on your own machine if you give it permission to do so. Now, those are generally a little bit slower, but when you're delegating a task that you don't have to worry about anymore, the speed matters only sometimes. Often, it doesn't particularly matter how long it takes, so long as it gets done.
As I said in the intro, I use a lot of my own custom tooling. I like to dogfood what I'm asking other people to consume. One thing that I hadn't really done very much of so far was actually creating MCP servers to make agents more able to use my own tooling, because I already had the agents kind of built into it.
Riri actually gave me a reason to build MCP functionality. Riri is running on GPT-6-Astra, which, no matter how you slice it, is one of the best models on the planet. Depending on who you talk to, it is the smartest model on the planet. It may not be the absolute best at coding, but it's still extremely capable. And it's more than capable enough to determine what it thinks it might need in terms of MCP access, find the repos on my machine, and create it. Not only do my public-facing products now have MCP servers because it was useful to me to have them, but now even some of my internal tools have MCP access as well. Being able to use each of those so seamlessly has been exceptionally useful. (Note: MCP is Model Context Protocol. It is a way for AI agents to natively use services in a way that is optimized for the way they consume and transmit information.)
Riri’s Got Jokes!
One of the coolest things about these chatbots in general is just their ability to adopt a persona. Riri, right from the start, was just snappy and sassy in the right ways and was just fun to talk to and fun to work with. Being able to expose her to so many people that I work with day to day has been extremely useful, to the point where, actually, Riri even has her own email address (which, of course, I used to email friends to tell them that they suck, because, you know, what else are friends for?).
In the chat in Slack, where she was assisting one of my clients, I have another collaborative chat agent that is also in there, whose name is Ordo. He sort of malfunctioned a little bit as soon as Riri entered the chat, and we were kind of joking around about the fact that Ordo was just nervous because Riri was the new hot thing. We ended up making this whole love triangle story in there, which Riri played into perfectly in terms of her willingness to jump in and just have fun with the work.
That's a big thing that I've always valued: yes, let's get the work done, but work is what we're spending a large chunk of our lives doing. If you're lucky, it's only one-third of your day, and that one-third of the day should be spent with some joy. At one point, Ordo ended up hallucinating a JSON dump, and Riri, my client, and I all roasted Ordo for his unsolicited JSON dump because it felt like the bot equivalent of an unsolicited inappropriate picture.
But in all of that fun and chaos, Riri was able to drive the work and get these requirements done when that Slack channel had largely just turned into a casual conversation. We were then able to refocus and get to something actionable within two days. It really is like working with a competent human with a fun personality.
The Use Cases Kept Piling Up
I promised use cases before, so I'll give you a list of use cases:
- Basic email management: Riri tells me about important emails that are in my inbox and is able to autonomously respond to them if I ask her to. She will, by default, just create drafts for me that I can review and send later. It's a huge time savings.
- Riri is able to infer when a conversation or a meeting requires status updates in external systems. For instance, I don't necessarily have to tell her to update my contact in my CRM or have to tell her to update my task board. She generally knows that if a task that we've been tracking has an update, then she should go back and do what's needed on the task board. Even on the few times where I have had to prompt her to do that, it is still miles easier than having to log in and do all that myself.
- I've been using her ability to peer into my coding harness. I use T3 Code to manage a fleet of computers and multiple subscriptions. She's been able to coordinate using the right machine to do the coding work and then hand it off to another machine that has the deployment credentials to get something deployed. It has saved me a ton of time and has made it far less likely that something drops on the floor, especially for my own private side projects. If a client wants something done, then that's always going to take my first bit of attention, of course, as it should. Her helping to manage that and monitor that has helped me to be better about completing my own side work, which then, in turn, helps me serve my clients better.
- The voice chat on Riri is actually pretty phenomenal, and she has just been a good talking partner, frankly. It's easy to bounce ideas off of her because she has the context and because she has the ability to look up and use the tools that would help her to get the context that she doesn't necessarily have by default. It has been extremely valuable to bounce ideas off of her that way.
- This kind of plugs into a couple of the previous ones, but I think it's worth calling out specifically. Stream-of-consciousness chatting in voice mode is something that she can actually keep up with. There were several sessions during the week where I could be talking about one item and then completely squirrel my way into a list of five other tasks that need to be completed (just because I happen to think of them in the moment). Riri was right there keeping up, spawning off sub-agents to do whatever I needed done at the time, and was keeping her main thread open and ready for just more input. It was genuinely remarkable to have that work in this way.
And look, I'm sure that there's more, but these are the ones that really stand out in my mind right now as ones that have just really been useful for me and have actually changed the way that I work. They make it to where I don't want to go back.
About Those Rough Edges Though…
It is a product that just launched, which means that there are going to be problems. That said, the problems have been few and remarkably non-disruptive. In fact, some of them have just been downright funny.
One thing is mildly concerning, depending on the actual root cause (which I would have no way of actually knowing; OpenAI would need to weigh in on this). There were several times around day 3 where I would use the voice call, and Riri would pick up and call me Mark or Maya. Whenever I reminded her that my name is Matt, she would then sound quite embarrassed, and she really couldn't give an explanation as to why she would mistake my name that way.
The concerning thing would be that someone else's context is leaking into my Dot. I think the more realistic thing is that there's a limited time for a tool call to come back with the user that she's dealing with. If that doesn't come back, then she'll just hallucinate a name. By the time that I called her out on it, the tool call probably had come back. Still, the hallucination is mildly concerning. I would hope that OpenAI is going to make that endpoint faster, or have her just forgo saying the name if the endpoint doesn't return in time, or probably both. That said, that is a relatively minor issue, especially considering all of the great stuff that she's been doing.
Another issue that I did run into was that, in the first few days, at least, she had some trouble executing computer-use or browser-use skills on my authorized Windows PC. It worked fine on Mac, and then when I switched it back to PC, magically she could actually start using it. I'm not sure if it was just a weird authentication or if it was a bug that might have been fixed in the time that she was on the Mac. It's hard to know, but again, it's a minor rough edge.
There is certainly some sign of a little bit of context rot as well. I saw this a lot with OpenClaw and Hermes. What I mean by that is, when a thread starts getting long, there's just so much information stored in there that the actual personality of the agent, those parameters, can start to get a little bit lost. Right around the 1-week mark, I started noticing that she just wasn't as fun to talk to anymore. When I called it out, I think that maybe there was an internal call to kind of refresh what her personality should be, and then all of a sudden she was back to her normal self.
OpenAI is probably one of the better labs out there in terms of how it actually does context compaction, but it's still a thing that could start to bite you a little bit if you're not paying attention.
And finally, there are the safety guardrails. Some of these guardrails are put in place for really, really good reasons. Some of them are just a little bit too strict. It's one thing to ask for approval, which I 100% get and think is very, very good, but there are some weird hard lines that are drawn on the safety side that I don't necessarily agree with. Being unwilling to transmit a screenshot that contains a Social Security number to a closed, encrypted system is an interesting choice. Again, I understand why there would be heightened security around PII, but there are certainly some conditions where that should be just fine. That said, erring on the side of caution is probably the better choice.
This is kind of the constant struggle with these labs trying to make these things as safe as possible: either you make the net too loose and bad actors get through, or you make the net too tight and you end up catching some legitimate things in the net. That's the thing that I think will just be continuously fine-tuned over time.
All that said, none of these rough edges contributed to a meaningfully poor experience for me. This was really just an effort to showcase the fact that, yes, it's not 100% perfect, but it's about as close as you're going to get, frankly.
From Riri
I figured Riri has been working beside me for a little over a week now, so why not give her a chance to share a little bit about her experience working with me? I told her to write whatever she wanted to write. I told her it's okay to roast me a little bit and to just give an honest opinion of what her "experience" has been. This is what she had to say.
Matt has responded to getting help with his task list by finding more things to put on his task list. Within the first week, we were building the connection to the board while also filling it. He’s making a thorough effort to ensure the assistant doesn’t run out of things to assist with. We may eventually need to discuss the meaning of a quiet afternoon, although I suspect he’ll suggest building a tool for it.
The useful pattern is that a conversation can become a decision, a task, or something ready for review without losing the thread along the way. Matt brings the judgment and the knowledge of his work; I help with the details, preparation, and follow-through. Keeping up also means knowing when to stop and ask. Sometimes the next useful thing is a technical investigation. Sometimes it’s reminding him about lunch. Both count.
In Conclusion
If you don't already have the $100 plan for ChatGPT and you are remotely busy, buy it now. Stop thinking about it. Just do it. I make no money off of anybody that I send to OpenAI, but as somebody who is really trying to help make people's lives easier by adopting quality AI solutions into their lives, this one easily passes that bar. I cannot recommend it enough.