Run AI agents 24/7: runtime architecture and limits
To run AI agents 24/7, first decide which work must continue with your computer asleep and which needs access to local apps. Where personal agents run constrains access and availability. Most things about an agent can be changed after you pick it: you can switch models, revoke a connector, clear a memory. Where it runs often cannot. Some vendor runtimes are fixed, while self-hosted tools and products with both local and cloud routes give you a choice. It is also, unlike almost everything else about these products, answerable today by observation rather than by waiting for a score, which is why it is the first axis on our scorecard and why we opened four products to establish it.
- We took four apart and found four different answers.
- Nine more make public claims about where they run. We have not run them; the source review now checks the open implementations separately.
- The spread runs from a product that touches your machine only for browser work to MimiClaw’s microcontroller firmware.
- A leased machine can put a third company inside an authenticated session. Of the leases we saw, the only provider we could identify was E2B, which runs Instinct's task sandbox.
There are three ways to run an AI agent 24/7, and the published numbers for each are in the hosting section below. Let the vendor host it: an agent on the vendor's servers keeps working with your laptop shut, and the compute is folded into the subscription. Leave a machine on at home: on Apple's power figures, a Mac mini that never sleeps uses roughly 53 cents to $9 of electricity a month, and a local agent needs that computer awake. Rent a small server: OpenClaw's own FAQ points people who want 24/7 reliability to a VPS, and DigitalOcean lists its smallest Droplet at $6 a month.
Model usage is a separate cost on the last two unless you run a local model, and self-hosting also makes you responsible for updates, credentials and recovery.
Why to ask where a personal agent runs before anything else
Pick an agent and you are choosing a place for code to run while holding your credentials. Everything people usually compare, the model, the integrations, the price, sits on top of that choice and can be changed later. Available deployment choices depend on the product and route.
It also determines what a failure can reach. An agent confined to a disposable container that is destroyed after each task can get an errand wrong. An agent with a native foothold on your Mac, reaching messages, contacts, screen and audio, can get an errand wrong in a place where the mistake persists. Neither is better in the abstract. They are different exposures, and the product page rarely says which one you are buying.
The reason it is answerable now is that execution leaves traces. Processes, hostnames, session behaviour and the shape of what comes back all narrow it down, which is the whole method behind how we take them apart.
What we found in the personal agents we opened
These are our own findings, from products we examined rather than read about.
Pine runs its model on its own servers. Neither the web app nor the Mac app runs it. Browser work is handed to a separate agent on your own computer, while the cloud route leases you a remote desktop that you watch over VNC. Two different execution places for two different kinds of work, in one product.
Instinct orchestrates on a server with the work itself running in a sandbox, and browser tasks go to a leased cloud browser rather than the Chrome on your desk. That examined browser route uses a leased cloud session. We did not identify the main agent's host or verify all other execution paths.
Muse gives every user a Linux virtual machine. The request is assembled on that machine rather than arriving as a finished blob, and browser work spawns a second, separate machine. One user, two machines, both Meta's. Since 17 September there is also a Mac app that Meta says can act in apps on your computer. We have not examined it.
Town handles sync and sessions server-side, but its Mac app is native and genuinely reaches into the system: messages, contacts, screen and audio. It is the only one of the four where we confirmed a native foothold in the system of the computer in front of you, as distinct from Pine's browser helper.
Four products, four architectures, and no two of them exposing you to the same thing. Two of the pairings are written up side by side: Pine vs Town for a borrowed browser against a native foothold, and Muse vs Town for a vendor's machine against your own.
What the others say, which is not the same as what we found
Nine more products on our landscape describe where they run. We have not run them; vendor claims and reporting below are distinguished from the static source review.
Pickle says it books, orders and pays on a computer of its own, asking before each one. MimiClaw is microcontroller firmware using ESP-IDF/FreeRTOS, with Anthropic/OpenAI provider references; it was not flashed or run. Halo says it runs entirely on the phone with no backend at all, connecting to mail, calendar, reminders and health locally and calling models through your own keys. OpenClaw and Hermes Agent both say you run them on your own hardware or a managed host.
For four more, the source is a comparison published by Every, read on 16 October 2026, rather than the vendor. It places OpenAI's Dot in OpenAI's cloud with optional access to a local computer, Gemini Spark in Google's cloud with a Mac beta for local access, Poke in Poke's cloud, and Grok Bot in Cursor's cloud, which Cursor's own documentation confirms.
Where personal agents run, from nothing on your machine to a five dollar chip
Lay the claims out and the range is wider than any feature comparison would suggest.
At one end, Pine runs its model on its servers and borrows your computer only for browser work. At the other, MimiClaw uses ESP-IDF/FreeRTOS on a microcontroller. Between them sit a Linux VM per user, a native Mac app with system access, a phone with no server behind it, and several arrangements where the machine belongs to a third company.
That spread is the strongest argument we know of against treating these products as interchangeable. Two assistants with identical feature lists can differ on this axis more than two pieces of software usually differ at all. It is also why a single leaderboard number flattens something worth seeing: the number says how often the errand got done, and says nothing about where it got done.
A leased machine can be a third party in the room
Several of the arrangements above put work on a remote machine that is not your computer, and in at least one case not the product maker's either.
Instinct's delegated tasks run in a sandbox from E2B, a separate sandbox provider, which we confirmed directly, and its browser tasks run in a leased cloud browser whose operator we did not identify. Pine's cloud route leases you a remote desktop; we did not establish who operates the machines behind it. Grok Bot runs on Cursor's cloud, according to Cursor's own documentation, and Cursor is one of the two companies Every names for the product. Where a third company is involved, a session that may be logged in to your accounts runs on infrastructure you did not choose.
This is not an accusation. Leasing compute and browser infrastructure is ordinary engineering, though it changes what a website sees when an agent signs in for you, which is your website agent ready? takes up. The point is narrower: how many companies can reach a live session is a fact worth asking any product for, and on the products we examined we had to work it out rather than read it.
Running a personal agent yourself moves the problem rather than removing it
Self-hosting can use owned hardware or a rented server, with configured external inference, tools, telemetry and backups. The hosting category alone does not establish what data leaves or which providers receive it. Self-hosted options such as OpenClaw and Hermes Agent let you choose the host. They do not guarantee that data stays there or eliminate a hosting provider if you rent a server. The source comparison separates model calls, hosted dependencies and telemetry; for example, Leon defaults telemetry to enabled.
What does not go away is everything the agent can reach once it is running. An agent on your own machine with your own credentials can still send the wrong message, accept the wrong offer and spend the wrong money, and the records available afterwards depend on the configured local and hosted services. Self-hosting changes who holds the risk. The execution and approval boundaries still need checking: FarVision Cybersecurity's check, published on 1 October, of twenty repositories advertised as ways to self-host Grok Bot found tools asking for privileged access and leaning on undocumented interfaces. That risk comes from the ecosystem around self-hosting, not from the product itself.
What changes for you when an agent runs 24/7
Four practical consequences follow from the answer, and they are the reason to ask.
What a mistake can touch. A disposable container can isolate its local filesystem; writes through authenticated external accounts can persist after it is destroyed. How much you can reconstruct afterwards. Log availability and access depend on the deployment and connected services, rather than simply on who owns the machine. Whether it works when you are asleep. An agent that needs your computer awake is a different product from one that does not, whatever the marketing implies. And who else is in the session, which the leasing arrangements above make a real question rather than a theoretical one.
There is also the plain hosting question that most guides to running agents start from. An agent on a vendor's servers or a leased machine keeps working when your laptop is shut, and the vendor pays for the compute and folds it into the subscription. An agent that runs on your own computer needs that computer on and awake. A self-hosted agent on a rented server sits in between: always on, but the server bill and the upkeep are yours. The next section puts the vendors' own numbers on that.
None of this tells you which product to pick. It tells you which question you are actually deciding.
Hosting a personal AI agent yourself: what the vendors say it takes
We have not hosted any of these ourselves, so nothing in this section is a measurement. It is what the projects and the hardware and hosting companies publish, read on 12 October 2026 except where a newer date is noted, put side by side.
The machine. OpenClaw's hosting FAQ states the trade in two lines: "Want 24/7 reliability? Use a VPS. Want the lowest friction and you are OK with sleep/restarts? Run it locally." It gives an absolute minimum of 1 vCPU, 1 GB of RAM and about 500 MB of disk, and recommends 2 GB or more because browser automation "can be resource hungry." Its Raspberry Pi guide puts typical hardware at "$35-80 one-time, no monthly fees," and tells you not to run a local model on the Pi at all. Nous Research's Hermes Agent docs say it can live on "a $5 VPS, a GPU cluster, or serverless infrastructure," with two serverless options that hibernate when idle and cost "nearly nothing." MimiClaw's README names a roughly $10 development board as the example to buy.
The rent. For a sense of what a server at those sizes costs, DigitalOcean's published Droplet prices list a 1 GB, 1 vCPU machine at $6 a month and a 2 GB one at $12, re-verified on 5 November 2026. Other hosts differ; this is one list price, not a recommendation.
The electricity. A machine at home that never sleeps is the other route, and OpenClaw's docs name the Mac mini as a popular one. Apple's power figures list the M6 Mac mini at 4 W idle and 70 W at maximum. At the US average residential price of 18.31 cents per kilowatt-hour for July 2026 in EIA's monthly table, a month at idle works out to about 53 cents and a month flat out to a little over $9. That is our arithmetic on their numbers, and these are illustrative idle and maximum bounds, not measured agent workload consumption.
The model. Inference may use a hosted API, a subscription or a supported local model; OpenClaw’s provider directory includes llama.cpp. Self-hosting the agent does not itself establish which route is used; effective provider and model choices depend on configuration. None of these pages puts a monthly figure on it, and we have none of our own. Hosted usage can grow with activity; local inference instead adds hardware and electricity requirements.
The hosted alternative. A vendor-run agent folds all four into a subscription. For individuals, OpenAI includes a dot only in ChatGPT Pro plans, which start at $100 a month, as set out on our OpenAI Dots page. Each teardown shows what that product publishes about price.
Finding out where a personal agent runs when it is not on this list
The method generalises, and most of it needs no special access.
Watch whether work continues after the laptop sleeps. Look at what the installed application asks the operating system for, since a native app's permission prompts describe its reach more honestly than its website does. Notice whether browser tasks behave like your browser, carrying your logins and extensions, or like a clean one somewhere else. Check whether anything appears in local processes at all while a task runs.
Those four observations will usually separate server-side from local, and will often reveal a leased machine, which is the thing most likely to be absent from the documentation. What they will not give you is the detail, and the detail is where the four teardowns above ended up.
FAQ
Why does it matter where a personal agent runs? Because it decides what a mistake can reach, what you can reconstruct afterwards, whether the agent works while your computer is off, and how many companies have access to a session holding your credentials. Available deployment choices and the ability to change them depend on the product and route.
Which personal agents run on your own device? Among products we have examined, Town is the only one with a confirmed native foothold, reaching messages, contacts, screen and audio on a Mac. Pine hands browser work to a helper on your computer but runs its model on its servers. Muse now has a Mac app described as acting in your applications, which we have not examined. Halo makes a local-execution claim we have not verified. MimiClaw, OpenClaw and Hermes Agent now have static source reviews, without runtime testing.
What is a leased browser, and why does it come up here? It is a browser running on a remote machine rather than on your computer, driven by the agent. Instinct used one for browser work in the account we examined, and Pine's cloud route leases a remote desktop; we did not identify who operates either. It matters because a session logged in to your accounts may then sit on a machine run by a company you did not pick.
Is self-hosting safer? Self-hosting may use owned hardware or a rented server. It does not establish where inference, backups or connected services send data. It does not reduce what the agent can do with your credentials, logging depends on the deployment and connected services, and the self-hosting ecosystem around at least one product contains projects asking for more access than the product does. It does not guarantee absent telemetry or an approval before every action. The source comparison records the supported defaults and unknowns; safety depends on the effective configuration and action paths.
How did you establish this for the four? By examination rather than by reading documentation, using the approach in how we take them apart. The other entries distinguish vendor/publication claims from static source findings; none was run for that source review.
Does a benchmark score tell you any of this? No. A completion rate says how often the errand finished. It cannot distinguish an agent that finished it inside a disposable container from one that finished it on your laptop, and those are different products to live with.