Step-by-step guides for running AI agents on their own computers: coding agents, secrets and egress, always-on assistants, teams and the EU.
Common questions about running AI agents on their own computers.
Yes. Sanlabs' forward-deployed engineers build AI agent use cases with your team on Sanbox, with experience in aviation, pharma, supply chain and logistics, and internal admin.
Don't give a coding agent a shell on production. Give it a Computer with staging-only credentials, deny-by-default egress and a snapshot, and review its diff before you deploy.
Create one template for your product and one Sanbox Computer per user, keyed by your user ID. Stop it when they leave, resume it when they return, and use warm pools for fast first starts.
Build one team template with a network policy and secret proxies, then start one persistent Sanbox Computer per person with a stable external run ID. Enterprise adds SSO.
Run Claude Code on a machine that stays on, in a terminal session that survives disconnects, and reach it from your phone with Remote Control or a browser terminal.
Give each task its own machine. sanbox batch runs every task in a JSON file in its own isolated Sanbox run, so tasks never share files, ports or databases.
To the provider and endpoint you configure. Claude via Bedrock or Google Cloud, OpenAI or Azure OpenAI with EU residency, and Mistral's EU endpoint keep inference in Europe; the exact scope differs by route.
Keep the key off the agent's machine. A Secret Proxy stores it and adds it to matching HTTPS requests on the way out, so the agent can call the API but never read the key.
Install Claude Code on a cloud machine whose disk outlives each session. A Sanbox Computer keeps your repository, installed packages and Claude Code settings across stop and resume.
Create a Sanbox Hermes Computer template with the email and Telegram channels, then start one always-on run with an AgentMail address and your Telegram bot. Memory and skills stay on its disk.
Use an ephemeral sandbox for one-shot code execution. Use a persistent computer when the agent must keep files, packages, browser logins or services between sessions.
Install the Codex CLI on a Linux machine that stays on, sign in with a device code or API key, and run it in a terminal session that survives disconnects, for example on a Sanbox Computer.
On Sanbox, a stop keeps the disk: /workspace, installed packages, Docker images and the browser profile. RAM, running processes, connections and exec commands end.
What an agent computer is, how it differs from a sandbox or a VPS, and what it keeps between sessions.
An agent computer is a persistent, isolated machine that an AI agent owns. It differs from a short-lived code sandbox and from a VPS that you run and maintain yourself.
Use an ephemeral sandbox for one-shot code execution. Use a persistent computer when the agent must keep files, packages, browser logins or services between sessions.
On Sanbox, a stop keeps the disk: /workspace, installed packages, Docker images and the browser profile. RAM, running processes, connections and exec commands end.
Containers share the host kernel. MicroVMs such as Firecracker give each workload its own kernel behind hardware virtualisation. Full VMs do too, with more weight.
Run Claude Code, Codex and OpenCode on a computer that stays on when your laptop doesn't.
Run Claude Code on a machine that stays on, in a terminal session that survives disconnects, and reach it from your phone with Remote Control or a browser terminal.
Install Claude Code on a cloud machine whose disk outlives each session. A Sanbox Computer keeps your repository, installed packages and Claude Code settings across stop and resume.
Install the Codex CLI on a Linux machine that stays on, sign in with a device code or API key, and run it in a terminal session that survives disconnects, for example on a Sanbox Computer.
Run opencode serve on a machine that stays on and connect with the OpenCode SDK. A Sanbox OpenCode Computer runs the server privately and gives your app a short-lived URL and token.
Claude Code: paste the sign-in code back, use a claude setup-token token, or an API key. Codex: codex login --device-auth or an API key. Plus where each stores the login.
Run many agent tasks at once, start from a known state, and keep long jobs going.
Give each task its own machine. sanbox batch runs every task in a JSON file in its own isolated Sanbox run, so tasks never share files, ports or databases.
Get one Computer to the starting point, stop it, and fork it three times. Each fork is an independent copy with the same files and packages, so three ideas can run without touching the original.
Use a Sanbox Docker Computer: a microVM with its own Docker Engine, Compose and Buildx. Allow your registries, copy the project in, and run docker compose through sanbox exec.
Allow the dev server's port on the Computer's template, then run sanbox ports forward to open it on your own machine. HTTP and SSE work; WebSockets, and so hot reload, don't.
Sanbox starts your agent at the times you choose, with no laptop or server of yours involved. Scheduled runs are enabled per organisation on request; today, any scheduler can call the CLI.
Keep credentials away from agents, limit where they can connect, and contain mistakes.
Don't give a coding agent a shell on production. Give it a Computer with staging-only credentials, deny-by-default egress and a snapshot, and review its diff before you deploy.
Keep the key off the agent's machine. A Secret Proxy stores it and adds it to matching HTTPS requests on the way out, so the agent can call the API but never read the key.
Use a deny-by-default egress allowlist. On Sanbox, a template's network policy lists the hosts and ports its Computers may reach; a blocked request names the host.
Run it in a disposable Computer: snapshot the stopped Computer, run Claude Code with --dangerously-skip-permissions, review the diff, and start fresh from the snapshot if needed.
Host Hermes Agent, OpenClaw and other assistants that run around the clock.
Over 12 months a small VPS is cheapest and a Mac mini pays off only over years or for macOS. A managed agent computer adds isolation, snapshots and no server to run.
Create a Sanbox Hermes Computer template with the email and Telegram channels, then start one always-on run with an AgentMail address and your Telegram bot. Memory and skills stay on its disk.
OpenClaw's docs say to use a VPS for 24/7 reliability and a Mac only for iMessage or macOS tools. A one-click OpenClaw Computer on Sanbox is enabled on request.
Sanbox gives an agent its own phone number, so people can reach it by text message. It's enabled per organisation on request. Today, Hermes Computers take email and Telegram.
Give agents a browser that stays logged in, and watch or take over when they need you.
Keep the browser profile on a disk that outlives each session and log in once yourself. A Sanbox Browser Computer keeps cookies and signed-in sessions across stop and resume.
Yes. A Sanbox Browser Computer shows the agent's chat beside its live browser in the console. You can watch each step and take control to complete a sign-in or verification step yourself, then hand back.
Give every employee an agent computer, let agents produce real work files, and get your use case built with us.
Yes. Sanlabs' forward-deployed engineers build AI agent use cases with your team on Sanbox, with experience in aviation, pharma, supply chain and logistics, and internal admin.
Build one team template with a network policy and secret proxies, then start one persistent Sanbox Computer per person with a stable external run ID. Enterprise adds SSO.
Yes. A Sanbox OpenCode Computer ships LibreOffice Writer and Calc, pandoc, python-docx, openpyxl, xlsxwriter and reportlab, so an agent can write .docx, .xlsx and .pdf files you download.
Sanbox connects agent Computers to Slack, Microsoft Teams and WhatsApp, enabled per organisation on request. Email and Telegram work today through Hermes Computers.
Sanbox shows and caps spend per Computer and per team, enabled per organisation on request. Today, hosting is agreed per organisation and your model provider's limits apply.
Give each of your users a persistent agent computer, with fast starts and your own models.
Create one template for your product and one Sanbox Computer per user, keyed by your user ID. Stop it when they leave, resume it when they return, and use warm pools for fast first starts.
Prepare a Computer once, snapshot it, and keep a warm pool of ready copies, so a new Computer can start in under a second. Snapshot templates let you promote or roll back.
Add the endpoint as a Custom OpenAI Proxy with sanbox model-providers create, then build templates on it. Any OpenAI-compatible URL works, such as an Azure OpenAI EU Data Zone deployment.
Connect your Supabase project to Sanbox once, then create a Hermes Computer per user with their Supabase Auth ID. The user approves access, and every read runs as them under your RLS policies.
The Sanbox MCP server lets Claude, ChatGPT, Cursor and other MCP clients start and use Sanbox Computers as tools. It's enabled on request; today, agents with a shell use the Sanbox CLI.
Call agent computers from n8n, Make, Zapier and other automation tools.
A native Sanbox node for n8n is enabled per organisation on request. Today, self-hosted n8n runs the Sanbox CLI from its Execute Command node, and Make or Zapier can call that workflow.
Where agent data and model calls go, and which contracts you need in the EU.
To the provider and endpoint you configure. Claude via Bedrock or Google Cloud, OpenAI or Azure OpenAI with EU residency, and Mistral's EU endpoint keep inference in Europe; the exact scope differs by route.
Usually, yes. If an AI agent processes personal data on a vendor's infrastructure or model, Article 28 GDPR normally requires a DPA with each such vendor, including the model provider.
Fixes for common problems when agents run on Sanbox, from blocked websites and package registries to proxy and certificate errors.
Sanbox Computers block outbound traffic by default. A blocked request names the host it couldn't reach: add that host and port to the template's network policy and start a new run.
How Sanbox compares with other ways to run agents, including where the others are stronger.
Eight ways to run AI agent sandboxes in the EU, compared by contracting company, location, persistence, DPA and price, including where each is stronger.
E2B gives agents SDK-driven Firecracker sandboxes, billed per second. Sanbox gives each agent a persistent Computer that stays on until stopped, in EU data centres, for a flat plan.
Daytona offers container, Linux VM, Windows, macOS and GPU sandboxes with five SDKs and per-second pricing. Sanbox gives each agent a persistent microVM Computer in EU data centres for a flat plan.
Docker Sandboxes run coding agents in a microVM on your machine for free, or in Docker's cloud with a one-hour default expiry. A persistent cloud computer runs until you stop it.
Claude Code on the web runs each session in a fresh Anthropic VM cloned from GitHub, included in your Claude plan. Your own cloud computer keeps its disk, services and tools, and runs any agent.
Ratgeber zu Forward-Deployed Engineering, EU, DSGVO und Auftragsverarbeitung.
Ja. Die Forward-Deployed Engineers der Sanlabs GmbH bauen KI-Agenten mit Ihrem Team auf Sanbox, mit Erfahrung in Luftfahrt, Pharma, Lieferkette und Logistik sowie interner Verwaltung.
Zu dem Anbieter und Endpunkt, den Sie konfigurieren. Claude über Bedrock oder Google Cloud, OpenAI oder Azure OpenAI mit EU-Datenresidenz und Mistrals EU-Endpunkt halten die Inferenz in Europa, je nach Weg mit anderem Umfang.
Meist ja. Verarbeiten KI-Agenten personenbezogene Daten auf der Infrastruktur oder mit dem Modell eines Anbieters, verlangt Art. 28 DSGVO in der Regel einen AVV, auch mit dem Modellanbieter.