Use casesPersonal agent
Hermes is a personal AI agent you host yourself. It needs a small machine that stays on and keeps its files. Only one of the sixteen providers offers it in one click; most can run it if you can follow terminal instructions; a few are a poor fit for an assistant that never sleeps.
“I want a personal assistant that answers me on Telegram, remembers what I tell it and runs jobs on a schedule.”
Hermes Agent is an open-source (MIT) assistant from Nous Research. You talk to it from Telegram, Slack, WhatsApp, email and 20+ other apps, it keeps a memory of you between conversations, and it writes itself reusable skills from what it does.
The model (Claude, GPT, an open model) runs elsewhere and is billed by its own provider. What you host is the small program that holds your memory, your logins and your schedule, and that runs commands, browses and edits files for you. That program is what this page is about.
Its own README says to run it "on a $5 VPS, a GPU cluster, or serverless infrastructure that costs nearly nothing when idle". A sandbox provider is one of those choices: the difference is how easy it is to set up, whether it is allowed to stay on, and what it costs when it is waiting.
From its own documentation, each line linked to the page it comes from
At least 1 CPU and 1 GB of memory; 2 CPUs and 2 to 4 GB recommended, and at least 2 GB if it uses the browser. 2 vCPU / 4 GiB / 20 GiB leaves room for Chromium and a growing memory.
Yes, usually. By default it fetches Telegram messages by asking Telegram every few seconds, and the docs say plainly that "a polling bot can never sleep". It also runs your scheduled jobs every 60 seconds. Webhook mode lets it sleep on hosts that wake on an incoming request, but scheduled jobs then do not fire while it sleeps.
Everything in ~/.hermes: memory, past conversations, skills, scheduled jobs, API keys and chat logins. A provider that deletes the disk when the machine stops loses all of it.
A browser is optional and is the hungriest part. Docker is optional. A GPU is only needed if you run the model yourself.
Mostly waiting. Battleships' always-on assistant workload averages about 6% CPU, with peaks near 100% when it browses or runs a command. That is why providers that charge only for CPU actually used come out cheap here.
battleships.dev, built by the Boat team
Host Hermes on the machine. Hermes itself lives in the sandbox: it holds your memory, answers your messages and runs around the clock. This is what most people mean, and it needs a provider that lets a machine stay on and keeps its disk.
Keep Hermes where it is, send its commands to a sandbox. Hermes runs on your laptop or server, and only the shell commands it decides to run go to a cloud sandbox, so a mistake cannot damage your own machine. Hermes has this built in for Modal, Daytona and Vercel Sandbox, and Fly.io publishes the same for Sprites. It is cheap, because the sandbox only runs while a command does, but it is not hosting: your own machine still has to be on.
Easiest first. “Without a terminal” means someone who does not type commands can get it running from the provider's website.
| Provider | How | Without a terminal? | What it involves |
|---|---|---|---|
| Agent 37 | One click | Yes | Pick Hermes on the Personal plans and it is set up with a dashboard, task board, web terminal and chat channels. The free plan sleeps when idle; paid plans stay up between scheduled jobs.Agent 37 Personal ↗ |
| Boat | Preinstalled | No | Hermes is already installed on every sandbox. Create one with auto-stop turned off, run the setup wizard over SSH, and keep the gateway running as a service.Boat: always-on services ↗ |
| E2B | Official guide or template | No | E2B ships a ready-made hermes template you start from its command line. Sessions end after 24 hours on Pro and one hour on the free plan, so it suits Hermes doing tasks better than an assistant that never stops.E2B: Hermes Agent ↗ |
| Blaxel | By hand | No | Possible in a sandbox with a volume for its files, installed by hand. Blaxel's agent hosting caps a run at 15 minutes, so it is not the place for an always-on assistant.Blaxel: get started ↗ |
| boxd | By hand | No | boxd documents the personal-assistant pattern: create a machine, install the assistant, turn off auto-hibernate (otherwise it sleeps after four quiet hours) and let chat apps reach it.boxd: personal assistants ↗ |
| InstaBox | By hand | Partly | The dashboard creates a machine and opens a browser terminal and desktop; installing Hermes and keeping it running is typed by hand. No Hermes template.InstaBox: getting started ↗ |
| Lizard | By hand | Partly | The dashboard creates a sandbox and opens a terminal; Hermes is installed by hand. Lizard's one-click templates cover OpenClaw, not Hermes.Lizard: sandbox quickstart ↗ |
| Morph | By hand | No | Create a Devbox, install Hermes over SSH, and turn on wake-on-HTTP so it pauses between messages and wakes when one arrives.Morph: Devboxes ↗ |
| Runtime | By hand | Partly | Start a sandbox from the dashboard and you get a shell in the browser; install Hermes there and mark the sandbox persistent so it never idle-pauses. No Hermes guide.Runtime: getting started ↗ |
| Sandboxer | By hand | No | Nothing preinstalled and no guide. Create a sandbox with auto-stop off, run Hermes' installer over SSH, and keep the gateway running as a service.Sandboxer: long-running tasks ↗ |
| Tensorlake | By hand | No | Create a named sandbox with `--timeout 0` on Pro, install Hermes and keep it in a background process or under systemd. Suspend is the catch: Hermes polls Telegram every few seconds, so the sandbox has to stay running, and only Pro allows a session with no cap. Everything in `~/.hermes` sits on the ephemeral root disk, so a snapshot or a mounted volume is what saves it if the sandbox is ever terminated.Tensorlake: timeout and plan limits ↗ |
| Daytona | Runs its commands only | Partly | A built-in Hermes backend: Hermes on your own machine starts, stops and resumes a Daytona sandbox for its commands. Daytona's dashboard Playground opens a browser terminal if you want to host it there instead.Hermes: Daytona backend ↗ |
| Modal | Runs its commands only | No | A built-in Hermes backend for running its commands. Not an always-on host: Modal sandboxes end after 24 hours and cannot pause.Hermes: Modal backend ↗ |
| Sprites | Runs its commands only | No | Fly.io markets Sprites for agents like Hermes, but its only official Hermes integration is the backend mode. Hosting Hermes inside a Sprite means registering it as a service, because a Sprite pauses about 30 seconds after activity stops.Sprites: agent plugins ↗ |
| Vercel Sandbox | Runs its commands only | No | Vercel publishes an official guide, but it is for the backend mode: Hermes stays on your computer and each task runs in a sandbox.Vercel: Hermes ↗ |
| Cloudflare Sandbox | Not documented | No | Neither Cloudflare nor Hermes documents this. Cloudflare sandboxes also lose their disk when they stop, so memory would have to be saved elsewhere. |
Read from each provider's documentation on 29 October 2026. A provider not listed as one click can still be the right home; it means someone has to follow its instructions.
Three starting points for different people, each with its catch
The only provider of the sixteen that sets up Hermes in one click. Plans run from $3.99 a month (1 vCPU / 4 GB / 8 GB, bring your own model key) to $29.99 (2 vCPU / 8 GB / 20 GB, with a built-in browser).
Catch: The free plan sleeps after 15 idle minutes. These Personal plans are a separate product from the Agent 37 Cloud machines the calculator prices.
Hermes is preinstalled on a full virtual machine. The small size running all month is $13.14 of usage, inside the $20 monthly plan.
Catch: Auto-stop is on by default, so create the machine with it turned off. A stopped machine's disk is kept for only 14 days, and there is no free plan.
It charges for CPU only while it works, so an assistant that mostly waits costs about $25 a month at 2 vCPU / 4 GiB, and a persistent sandbox never idle-pauses.
Catch: No Hermes guide, one US region, and storage keeps billing while the machine is stopped.
Beyond these providers. Hermes' own docs point to an ordinary virtual private server ("a $5 VPS") or Docker on a machine you own. Neither isolates the agent the way a sandbox does, and neither is set up for you.
Compare these three head to head: features and cost side by side.
One always-on 2 vCPU / 4 GiB / 20 GiB machine that keeps its files. It waits for messages most of the time: about 6% CPU on average, with peaks near 100% when it browses or runs a command. Change any number; open a row for the arithmetic.
List prices only, excluding model tokens, egressData the machine sends out to the internet, such as downloads to a user or uploads to another service. Some providers charge per GB; most agents send very little., taxes and anything a provider does not publish. How the numbers are made.
Each page compares every provider, with a source for every row
Idle timers, waking and session limits — An always-available agent depends on three separate settings: a lifetime deadline, an idle timer, and a wake path. Providers bundle them differently, and the difference is most of your bill.
Lifecycle: stop, pause and sleep — Providers use seven words for four behaviours. The only question that matters is what is still running when the machine comes back.
Storage: disk, volumes and what persists — “Persistent” can mean the whole disk, one mounted folder, or fourteen days. It can also mean a separate volume that outlives the machine, which is a different product from the disk the machine boots with.
Reaching it: links, ports and who can open them — A machine nobody can reach is not an agent. There are four doors in — an HTTPS link, a raw TCP port, SSH, and the provider's API — and the important difference is who can open each one.
Keys and secrets — Three different things are called “secrets”: a value in an environment variable, a file on disk, and a placeholder that a proxy swaps for the real key. Only the third keeps the key away from the agent.
Billing models — The meter decides your bill more than the hourly rate does. Four shapes are in play: flat per size, per vCPU and per GB, a plan fee that returns as credit, and measured CPU.