Give an AI agent a shell, not your laptop
Three new templates give an AI agent its own machine on velixir: OpenClaw, Hermes Agent, and Steel Browser, a real Chrome for your agents to drive. Each is locked by default, starts fresh on every restart, and runs on your own model key.
The most useful AI agents stopped just talking a while ago. They run commands, install packages, write files and open websites. That is exactly what makes them worth trying, and exactly what makes you hesitate before letting one loose on your own laptop.
The usual alternative is an afternoon of setting up a server for it. So we made that part one click. Three new templates give an agent its own machine on velixir, locked down from the first boot, that starts fresh every time it restarts.
The three
OpenClaw is an open-source AI assistant with a web Control UI and an optional Telegram bot. It runs shell commands, writes files and reads web pages, all inside its own container. Set an Anthropic, OpenAI, Gemini, OpenRouter or xAI key and you are chatting with it in a few minutes. With an Anthropic key it starts on Claude Opus 5.5.
Hermes Agent is the self-improving agent from Nous Research. You talk to it on Telegram, Discord or Slack. It runs commands, searches the web, remembers what you tell it and writes itself new skills as it works. Bring a key from OpenRouter, Anthropic, OpenAI, Nous Portal or any OpenAI-compatible API.
Steel Browser is a real Chrome instance for your agents and scripts to drive. Open sessions over its REST API, connect Puppeteer or Playwright over a WebSocket, or turn any page into markdown, a screenshot or a PDF with one call. Its web UI shows the browser live while it works, and lets you take over the mouse and keyboard.
A restart is an undo button
These are sandboxes, on purpose. Whatever an agent installs, writes, learns or breaks is wiped when the app restarts or redeploys. Its settings are environment variables, so it comes back configured every time, just with a clean slate.
That makes them the right place to see what an agent can really do, to run experiments, and to hand off disposable jobs: research, a one-off script, a file conversion, a scraping run. It makes them the wrong place for an assistant that should remember you for months, and each template says so on its first line.
Locked by default
An agent with a shell, left open on the internet, is a shell for whoever finds it. So none of these starts open:
- OpenClaw needs its gateway token, and every browser has to be paired before it can do anything. Strangers who find your Telegram bot get a pairing code that only you can approve.
- Hermes Agent only answers the chat users you list, and refuses to start with a wildcard. It asks you in chat before deleting files or running anything similarly risky.
- Steel Browser checks your API key on every request and every connection, before anything reaches the browser. An open browser on the internet is a free proxy for anyone, and this one isn't.
If you don't set OpenClaw's token or Steel's key yourself, the template makes a strong one at startup and prints it once to the app's logs.
Give your coding agent a real browser
Steel works with the agents you already use. Deploy it, copy its key from the logs, and one command gives Claude Code browser tools that run in the sandboxed Chrome instead of yours:
claude mcp add steel-browser -- npx @playwright/mcp@latest \
--cdp-endpoint wss://your-app.velixir.run/ \
--cdp-header "steel-api-key: YOUR_KEY"
Cursor and other MCP clients take the same settings, and Python agents built on browser-use connect the same way. Your agent navigates, clicks and reads pages somewhere that is not your own browser, with none of your cookies or logins anywhere near it.
What you should know
They start empty after every restart. That is the point, but it means OpenClaw forgets paired browsers and Hermes forgets its memory. Changing an environment variable redeploys the app, so that resets it too.
Your prompts go to your model provider. The agents' machines and Steel's browser run in EU data centres. The model is whichever provider you bring a key for, and that provider bills the tokens.
OpenClaw and Hermes read the web but don't drive a browser here. To keep them small, their images leave Chromium out. Steel is the browser, for agents that speak its protocol.
Steel runs one session at a time. Starting a new session takes over the browser. For parallel work, deploy more than one.
They stay awake. Agents that listen on Telegram and browsers holding sessions can't sleep, so none of them runs on the Free plan. Hermes Agent fits in 512 MB on Hobby, €3.49 a month, which you can start without a card. OpenClaw and Steel need 2 GB on Standard, €12.95 a month. Adding a card unlocks it, and a first card comes with €14.99 of credit.
Web traffic only. Apps on velixir reach the internet on ports 80 and 443. Model APIs and websites work; email over SMTP and SSH to other machines don't.
Try one
Pick one on velixir.net/templates and click Deploy. Each template's page lists what it needs: nothing at all for Steel Browser, a model key for OpenClaw, and for Hermes a model key, a Telegram bot token and your Telegram user ID. From a terminal it is one command:
velixir deploy --template steel-browser
Swap in openclaw or hermes-agent for the others. Give an agent a machine of its own, and keep your
laptop for yourself.
Liked this post? velixir is hiring its first .NET customers.
Start free - your first month's on us