Self-hosted OpenClaw: review and managed alternative
A plain description of what running OpenClaw yourself involves, the facts you need before you commit to it, an honest review, and a side-by-side comparison with ControlClaw.
Updated November 7, 2026. OpenClaw details come from its public docs and GitHub. Spotted something out of date? Tell us.
What is self-hosted OpenClaw?
Self-hosted OpenClaw means you install OpenClaw on a machine you own or rent, and you run it yourself. OpenClaw is the open-source, always-on AI agent. It is free software under the MIT license, which lets anyone use and change it. It remembers things between sessions, connects to chat apps such as WhatsApp, Telegram, Slack and Discord, and acts through skills and tools. A background program called the Gateway keeps it online.
The machine can be a laptop, a Mac mini, a Raspberry Pi or a rented cloud server (a VPS) from any provider. The OpenClaw docs have guides for Hetzner, DigitalOcean, Oracle Cloud, GCP, Hostinger and others. You bring your own model key from OpenAI, Anthropic, Google or another provider, or sign in with a ChatGPT or Claude subscription. You pay them directly for what the agent uses.
In short: you get the whole agent, with no company in between. You also get the whole job: the server, the updates, the backups, the security settings and the repairs. That trade is what the rest of this page is about.
Self-hosted OpenClaw at a glance.
What the docs say you need, and what is left to you.
OpenClaw runs on macOS, Linux or Windows and needs Node.js, a program runtime, in version 24 or 26. One command installs it and starts a setup wizard. An option installs the Gateway as a background service so it survives a reboot. The docs give a bare minimum of 1 CPU and 1 GB of memory and recommend 2 GB or more. A server that size costs about $6 a month at DigitalOcean. Running it on a laptop costs nothing extra, but the agent stops when the lid closes or the network drops. That is why the docs recommend a machine that is always on.
Everything after the install is your configuration. By default the Gateway listens only on the machine it runs on. Reaching it from your phone means setting up a secure tunnel (SSH or Tailscale). Pairing codes and allow-lists you approve decide who may message the agent. The sandbox, which limits what the agent can touch, is off by default. Approvals for commands are off by default. The docs ask you to set up tool allow-lists, a read-only "reader" agent for untrusted content, and regular backups yourself. Updates are one command, openclaw update, and new versions ship several times a month. Support is the docs, GitHub and the community.
- Type
- Open-source agent you install and run yourself
- Agents
- Several separate agents in one Gateway; strict separation means separate Gateways
- Where it runs
- Any machine: your laptop, a Mac mini, a Raspberry Pi or a VPS from any provider
- Models
- Any provider by API key, or a ChatGPT or Claude subscription; local models via LM Studio
- Memory
- On your disk, in files and a database file;
openclaw backupexports it all - Setup
- One install command; channels, keys, the service, the firewall and backups are yours to set up
- Safety controls
- Pairing and allow-lists for senders; sandbox and command approvals exist but are off by default
- Team admin
- One trust boundary per Gateway; no admin console for a team
- Pricing
- Software is free; you pay for a server (from about $6/mo) and model usage
- Open source
- Yes, MIT license on GitHub
Self-hosted OpenClaw review: everything is yours, including the work
We run the competing product, so read this with that in mind. We have tried to be fair; where self-hosting is better, we say so.
Self-hosting is the cheapest and most direct way to run OpenClaw. The software costs nothing, a small server is a few dollars a month, and you pay your model provider only for what the agent uses. Nothing sits between you and the agent, no vendor and no plan limits. You can read the code, pick the model, change any setting and move the whole thing to another machine with one backup command. For a developer who likes to own their tools, that is the whole appeal.
The cost is time and attention. The docs are honest about this. The Gateway is private by default, but the safety features that matter for an agent with tools are off until you turn them on. The sandbox is off. Command approvals are off. There is no filter for prompt injection, where an email or web page carries hidden instructions for the agent. OpenClaw marks untrusted text so the model can see where it came from. After that, it is up to the model and your tool allow-lists. API keys, chat logins and conversation history live on the same machine the agent controls. The docs tell you to keep that machine patched, keep secrets out of the agent's reach, and back up regularly. Each of those is a task you do and keep doing, as new versions arrive two or three times a week.
The other limit is people. OpenClaw's trust model is one operator, or one team that trusts each other, per Gateway. There is no console where a manager sees what three agents did today. There is no approval queue a colleague can answer, and no allow-list for whom the agent may email. You can build some of that with configuration and a chat channel, but it is not in the box. For one person that is fine. It becomes a problem the first time a second person needs to know what the agent sent, and to whom.
What self-hosting does well
- Free software; the only bills are the server and the model
- Full control: any machine, any model, any setting
- Unmodified OpenClaw; every skill and channel works
- Memory, files and credentials stay on your own hardware
- Easy to leave: one backup command moves everything
Where it falls short
- Sandbox and command approvals are off until you turn them on
- No prompt-injection filter; untrusted content is marked, not screened
- Secrets live on the machine the agent controls
- No team console, approval queue or email allow-lists
- Updates, backups, patching and repairs are your job, every week
The right choice for one developer who wants their own agent and is happy to maintain it. Not a fit for a team, or for an agent that handles customer email, money or passwords, unless you will build and keep the guardrails yourself.
Same agent. Different amount of work.
Both run the open-source OpenClaw
Your skills, prompts and memory work on either. Neither changes the agent itself.
Doing it yourself versus a service around it
Self-hosting gives you OpenClaw and nothing else. ControlClaw adds the server, the AI Firewall, approvals, allow-lists and a team console around it.
Developers versus teams
Self-hosting fits a developer who wants to own the machine. ControlClaw fits a team putting agents on email, chat and calls.
Feature comparison: ControlClaw vs self-hosted OpenClaw
| Feature | ControlClaw | self-hosted OpenClaw |
|---|---|---|
| Basics | ||
| What you get | Managed OpenClaw agents plus AI Firewall, approvals and console | The OpenClaw software, free; the machine, the keys and the upkeep are yours |
| Open source | YesOpenClaw and ControlClaw, both on GitHub | YesOpenClaw, MIT license on GitHub |
| Setup | Sixty seconds from a template; channels and keys wired for you | One install command; you set up channels, keys, the service and the firewall |
| Servers | Hetzner or AWS in EU, US, UK, Singapore, or your own cloud | Any machine or VPS you choose |
| Safety | ||
| Secrets | YesHeld by the Firewall, never on the agent | PartialStored on the machine the agent runs on |
| Prompt-injection screening | YesInbound mail, pages and files inspected first | PartialUntrusted text is marked, not screened; a reader agent is yours to set up |
| Approvals | YesOne queue, full context, audit trail | PartialApprovals for shell commands, off by default; none for email or messages |
| Email and calendar allow-lists | YesBy domain, address or group | NoAllow-lists cover who may message the agent, not whom it may email |
| Spend and tool limits | YesPer agent, enforced outside the agent | PartialTool allow and deny lists per agent; spend limits sit with your model provider |
| Capabilities | ||
| Meetings and calls | YesJoins Meet, Zoom, Teams and phone calls | UnknownNot published |
| Multiple agents | YesFleet, templates, shared Brain | PartialSeveral agents in one Gateway; no fleet view or team console |
| Repairs | One-click repair session you confirm | openclaw doctor repairs config and the service when you run it |
| Backups | YesDaily and encrypted, included; kept 7 days and 4 weeks | openclaw backup, manual or a daily job you turn on; storage is yours |
| Commercial | ||
| Price | From $19/mo including a Small server | Software is free; a 1 GB VPS from about $6/mo, plus model usage |
| Trial | Fourteen days, no charge, card required | Free to run; no trial needed |
Where a row says "off by default" or "yours to set up", the feature exists in OpenClaw's settings but does nothing until you turn it on and maintain it.
Self-hosted OpenClaw or ControlClaw: match the job.
Pick self-hosted OpenClaw
You want to own the machine, not rent a service.
- You are one developer running one agent
- You already run servers and are happy to keep one patched
- The agent never touches customer email, money or passwords
- You want the lowest possible monthly cost
- You want no company between you and the agent
Pick ControlClaw
Your agents talk to people outside your team.
- Agents send email, join calls or hold passwords
- More than one person needs to see what they do
- You want a human in the loop before risky actions
- You run more than one agent, or plan to
- You would rather not do the updates, backups and repairs yourself
Self-hosted OpenClaw vs ControlClaw pricing.
ControlClaw bills the plan flat and passes the server through at list price. Self-hosting has no plan: you pay a server provider, a model provider, and your own time.
Self-hosted OpenClaw vs ControlClaw: common questions.
Is ControlClaw just hosted OpenClaw?
No. Hosting is part of it. ControlClaw also puts an AI Firewall in front of each agent, holds its secrets, runs an approval queue and gives a team one console for every agent. The agent inside is unmodified OpenClaw.
Can I move my self-hosted OpenClaw to ControlClaw?
Yes. Run openclaw backup create on your machine and import the workspace, memory and skills into a new agent. Channels are reconnected from the console.
Can I move back to self-hosting later?
Yes. Export the agent from ControlClaw and run OpenClaw anywhere. There is no proprietary format.
Which is cheaper?
Self-hosting. The software is free and a small server is about $6 a month, plus model usage. ControlClaw Solo starts at $19 a month including the server, firewall and approvals. The difference pays for the guardrails and the upkeep.
Can I bring my own server to ControlClaw instead?
Yes. Solo on your own cloud is $10/mo plus $5 per extra agent.
Sources (19)
- https://github.com/openclaw/openclaw
- https://docs.openclaw.ai/install
- https://docs.openclaw.ai/help/faq-first-run/providers-and-hosting
- https://docs.openclaw.ai/vps
- https://docs.openclaw.ai/install/digitalocean
- https://docs.openclaw.ai/install/hetzner
- https://www.digitalocean.com/pricing/droplets
- https://docs.openclaw.ai/gateway/security
- https://docs.openclaw.ai/gateway/security/trust-model
- https://docs.openclaw.ai/help/faq/security-and-access-control
- https://docs.openclaw.ai/gateway/security/prompt-injection
- https://docs.openclaw.ai/tools/exec-approvals
- https://docs.openclaw.ai/tools/exec
- https://docs.openclaw.ai/concepts/multi-agent
- https://docs.openclaw.ai/install/updating
- https://docs.openclaw.ai/releases
- https://docs.openclaw.ai/install/backups
- https://docs.openclaw.ai/cli/backup
- https://platform.claude.com/docs/en/about-claude/pricing