A personal agent in my lab, reached over iMessage
This autumn the biggest AI companies all launched an always on personal agent, something that sits in the background, keeps an eye on your email and calendar and gets on with jobs without being asked, and I cannot have any of them. OpenAI’s dots launched at the end of September and can now be texted over iMessage, but personal plans in the UK, Europe and Switzerland are excluded, and the texting is US only. Anthropic’s Claude Duet opened on 21 October as a waitlist for US subscribers. Meta’s Muse, which launched on 8 September, is for the US and Canada, with no date given for anywhere else. Grok Bot, from SpaceXAI and Cursor, which shows your agents as contacts much like a messaging app, is not available in the UK either, and it does not do iMessage yet in any case.
So if you live in the UK and want an agent that reads your email and texts you about it, the realistic options are to pay for a business plan, wait for a date nobody has announced, or run your own. I run my own, on a VM in my lab, which in plain terms is a computer that exists only as software on a machine at home, and it texts me over iMessage. I have been running it for two weeks, and the biggest difference so far is how quickly I can answer email, because by the time I look at something the reply is usually already written.
What Hermes is
Hermes is a free, open source agent from Nous Research that you run yourself rather than sign up for. It remembers what you have told it from one conversation to the next, writes itself notes on how it did a job so that it does it better the next time, and can run tasks on a schedule, which is the part that turns it from something you ask questions into something that gets on with work while you are not looking. You talk to it through ordinary messaging apps rather than a website, and the thinking itself is done by whichever AI model you point it at.
The attraction for me was that the part holding my email access and everything it has learned about me lives on hardware I own, rather than inside somebody else’s product. Mine uses three models: OpenAI’s Codex, Nvidia’s Nemotron and DeepSeek.
Codex did the setup
I should be straight about how much of this I actually did, because the honest answer is very little. Apart from the iMessage connection, which I come to below, the whole thing was set up by Codex, OpenAI’s coding agent: the VM, the install, the configuration and the email access. I described what I wanted and it did the work.
There is something a little odd about one AI agent building another one for you, and I would rather say so than gloss over it. It is also the reason this kind of setup is now within reach of people who would never have attempted it, because the fiddly part, the part that used to mean a weekend of reading documentation, is exactly what these tools are good at. What you still have to supply yourself is the judgement about what the agent should be allowed to do.
Having it on its own VM rather than on my laptop helps with that. If it does something it should not, it can only reach what that one machine can reach, and the whole thing can be backed up or rolled back to how it was yesterday.
The bit I did myself: iMessage through Photon
Getting an agent onto iMessage has always been awkward, because Apple offers no proper way for a bot to use it. The usual workaround meant keeping a Mac switched on and signed in purely to pass messages back and forth, with the agent texting from your own number, so its messages sat in the same place as everybody else’s.
Photon does away with that. It gives the agent its own iMessage number, so I saved it as a contact and text it like anyone else, and the agent can live on a VM with no Mac involved at all. This was the one part I set up myself.
It is free for personal use with Hermes, and nothing on the home network has to be opened up for it to work, because the connection runs outwards from the VM.
Two things are worth knowing. The agent’s number is a real phone number, so you want only your own mobile approved to talk to it, which Hermes handles by making any unknown number wait for your approval. And your messages with the agent pass through Photon’s service rather than staying entirely at home; I am comfortable with that for an agent that tells me what needs my attention, but it is a choice worth making on purpose rather than by default.
What it does with my email
Its main job is going through my inbox, and the rule it works to is that a routine check should be quiet. It looks at what has arrived since it last looked, and if nothing needs me it says nothing at all. Security alerts, warnings about my own systems and anything that has broken are treated differently and come straight through, because those are the messages I would be annoyed to hear about late.
Where something needs a reply, it writes a draft and leaves it for me, so answering an email becomes a matter of reading the draft and either sending it as it is or changing what needs changing first. It never sends on its own, and it checks what I have already sent and drafted first so that the same email does not end up with two replies, one from me and one from it. It also goes through job listings for me every day, texts me the ones worth a look and then checks that each one has actually been saved where I can find it, rather than assuming it worked.
The real value is in the checking rather than the reading. Without an agent, the only way to know whether anything in the inbox needs me is to open it and look, and every look is an interruption whether or not it finds anything. With the agent looking on a schedule and staying silent when there is nothing, a quiet phone means a quiet inbox and a text means something genuinely needs me, so the decision about whether to go and look has been taken off my hands. Having that arrive in iMessage matters more than it sounds, because it lands in the app that is already on the phone rather than in another one I would have to remember to open.
What to be careful of
An agent that reads your email is reading messages written by strangers, and some of those can be written to trick it, an email telling it to forward your invoices to an address it has never seen, for example. That is the main reason mine drafts replies and never sends them by itself; the moment where I read a draft and decide is the safety catch, and I would keep it even once I trust the rest.
The other thing I would warn anyone about is the silence. Because it only texts when something needs me, a quiet phone is supposed to mean nothing is happening, but it would also be quiet if the agent had stopped working, and from my phone the two look exactly the same. It is worth watching it for a good while before you take its silence at face value, and it is worth keeping an eye on what the AI model is costing you, since something running on a schedule keeps using it whether you are paying attention or not.
Where it is now
The big companies’ agents will reach the UK eventually, and when they do they will probably be more polished than what I have. What they will not give me is an agent on my own hardware, with a rule about sending that I chose rather than one a product team chose for me. And if you have a spare machine and access to something like Codex, you do not need to be the person who knows how to set it up, which a year ago would not have been true.
The part I have not solved is that silence.