OpenClaw is the open-source AI assistant that lives in your chat app and does real work on a machine you control. It reads your inbox, sends email, manages a calendar, runs commands and browses the web. Its own homepage calls it "the AI that really does things". A post quoted on that homepage credits "dozens of maintainers, and thousands of contributors" with building it in about six months.
So when someone on your team says "let's just run OpenClaw for everyone", two questions follow fast. Is OpenClaw safe? And if the software is free, what does it actually cost to run it for a team?
This post answers both from OpenClaw's own words. Everything about the project here comes from its homepage, its GitHub README, and its documentation on team setup, security, the trust model, prompt injection, secrets and storage and its FAQ, all read on 9 October 2026. We have not run OpenClaw on a team's real work for this post, so treat it as a careful reading of what the project publishes, not a hands-on test. We also sell a hosted AI employee called Sanaf, which is one of the options at the end. You should know that going in.
What is OpenClaw?
OpenClaw's FAQ puts it in one line: "an AI assistant you run on your own infrastructure, for yourself, or shared with your team."
You install it on a laptop, an office Mac or a small server. A program called the Gateway runs all the time on that machine and acts as the control room. You connect it to the chat apps you already use (WhatsApp, Telegram, Slack, Discord, Signal, iMessage, Google Chat, Microsoft Teams and more) and to whatever model you want to pay for or run locally. From then on you message it like a person and it does the work with the tools you allowed.
Three facts about it matter for everything below:
- It is free and open source. The license is MIT. The homepage says it plainly: "No subscription. No hosted tier. No token."
- It runs where you put it. State, memory and credentials "live on your hardware", as the README says. Nobody hosts it for you.
- A foundation looks after it. The OpenClaw Foundation, an independent 501(c)(3), employs the core team and signs releases.
Who owns OpenClaw?
This comes up a lot, because its creator now works at OpenAI. The FAQ answers it directly: "No. OpenClaw is stewarded by the OpenClaw Foundation, an independent 501(c)(3). OpenAI is one of several donors, and its creator works there. Donors do not own, control, or direct the project." The Foundation is funded by donations and is not venture-backed.
Is OpenClaw safe?
Short answer: it can be, and the project is unusually honest about what "safe" depends on. The longer answer depends on who is using it and what you let it touch.
The defaults are careful. OpenClaw's security page says the Gateway "binds to loopback" on a normal install, which means only the machine it runs on can reach it. Most chat channels answer a stranger's first message with a pairing code instead of doing what they ask. Group chats are allowlisted per room and the bot only replies when someone mentions it. There is also a built-in check you can run at any time:
openclaw security audit
That is a good start. The risk sits elsewhere, in three places the docs name themselves.
1. It acts with the access you give it
An agent that can run commands, read files and send email can do all of those things wrong. OpenClaw's docs say tools "run on the host for the main session unless you configure sandboxing", and that sandboxing is opt-in. If you turn it on, risky work runs in a container. If you don't, it runs on the machine.
Credentials also live on that machine. The secrets page says to "assume anything under ~/.openclaw/ may contain secrets or private data": the config file can hold tokens, a credentials folder holds channel logins, and a local database holds OAuth access and refresh tokens. The docs recommend full-disk encryption and a dedicated operating system user for the Gateway. Those are sensible steps, and someone on your team has to take them.
2. Prompt injection is still unsolved
Prompt injection is when text the agent reads, such as an email, a web page or a PDF, carries instructions that steer it into doing something you never asked for. OpenClaw's own prompt injection page says it plainly: "Prompt injection does not require public DMs." Even if only you can message the bot, "any untrusted content it reads" can carry an attack.
The same page gives the most useful numbers on the subject we have seen in a vendor doc. It reports that in a 2026 crowdsourced test of 272,000 attacks, success rates were 0.5 percent for Claude Opus 4.5 and 8.5 percent for Gemini 2.5 Pro, and it adds that adaptive human attackers still break well-scored models at rates above 80 percent (OpenClaw prompt injection guide). Its advice follows from that: use the strongest model you can for any agent with tools, keep exec, browser, web_fetch and web_search off unless you need them, and keep "hard enforcement" (tool policy, approvals, sandboxing, allowlists) "for anything whose blast radius you would not accept on a bad day."
So picking a cheaper model is also a security decision.
3. A shared gateway is one trust boundary
This is the part a team should read twice. OpenClaw's trust model says: "Everyone who can message a tool-enabled agent shares that agent's delegated tool authority." If five people can ask the bot to do things, the bot's access is effectively theirs too.
The docs are clear that this is fine "for a team whose members already trust each other", and equally clear about what it is not: "OpenClaw is not a hostile multi-tenant security boundary." Two defaults make the point concrete. Session visibility is set to all and agent-to-agent messaging is on, so an unsandboxed agent can "list, read, search, and message every agent's sessions, including other users' transcripts." The secrets page spells out the Slack case: if everyone in a workspace can message the bot, "any allowed sender can induce tool calls", and if the agent holds sensitive files or credentials, any of them "can potentially drive exfiltration."
None of this is hidden. It is written down, with config keys to tighten each piece. But it means the answer to "is OpenClaw safe for our team" depends on whether someone will read those pages, set those keys and keep them set.
How a team uses OpenClaw today
OpenClaw started out as a personal assistant, and a lot of comparison posts still describe it that way. The current docs go further. The README now says one Gateway runs "as a personal assistant on a laptop or as a shared team deployment; configuration is the only difference."
The team setup guide walks through it in five steps:
- Give the team access. The Gateway stays on loopback and teammates reach it through a tailnet with Tailscale Serve, an identity-aware proxy like Cloudflare Access, or a shared token for a small team.
- Connect the team chat, for example a Slack bot allowed in one channel that replies only when mentioned. Direct messages still need a pairing code the first time.
- Sign people in to the Control UI. Each person gets a profile, and identity-backed sign-in makes the history say who did what.
- Work in shared sessions. Every session records its creator, its current owner and everyone who prompted it, and the UI shows who is viewing and typing.
- Set limits per person. Named roles set which sessions someone can touch, which agents they can use, and whether their sessions must be sandboxed.
It also supports multiple agents on one Gateway, each with its own workspace, for projects that should not share memory. And it gives one rule for when to split: "Mutually untrusted users, customers, or organizations: separate gateways."
For a technical team that trusts itself, that is a serious, well-documented setup. What it asks for in return is a host that stays on, someone who owns the config, and a habit of running the audit.
Is OpenClaw free? What it really costs
The software is free. Running it is not, and the bill lands in three places.
Model usage comes first. You bring your own model keys, so every task is billed by Anthropic, OpenAI or whichever provider you picked, at their rates. The security docs push you toward the strongest model tier for any agent with tools, which is also the most expensive one. A cheaper model saves money and, by OpenClaw's own account, makes the agent easier to steer.
Then hosting. A team gateway needs a machine that stays on: a small cloud server, an office Mac or similar. Add the tailnet or proxy in front of it if you follow the guide.
Then people, which is the line nobody budgets for. Someone installs it, writes the config, sets the roles, keeps up with releases (the macOS app was on v2026.9.9 when we checked), runs openclaw security audit, rotates credentials, and answers for it when it does something odd. The docs even include an incident response page with steps to contain, rotate and audit after a suspected compromise. That is the work a hosted product's price is meant to cover.
So "OpenClaw pricing" really means model spend plus a server plus a few hours of a technical person's time every month. For one engineer who enjoys it, that is a great deal. For a team where nobody wants the job, it is easy to underestimate.
A short checklist before you share OpenClaw
If you are going ahead, these come straight from OpenClaw's own docs and are worth doing on day one:
- Run it on a dedicated machine, VM or container with full-disk encryption and its own operating system user.
- Don't sign that runtime into anyone's personal Google, Apple or password manager accounts.
- Keep the Gateway on loopback and give teammates access through Tailscale or an identity-aware proxy instead of a public port.
- Allowlist the channels it may answer in and require a mention.
- Turn on sandboxing for anyone who isn't an administrator, and keep
browser,web_fetchandweb_searchoff unless a job needs them. - Set
tools.sessions.visibilityandtools.agentToAgentdeliberately instead of leaving the defaults. - Use the strongest model you can afford for any agent with tools.
- Run
openclaw security auditafter every config change and fix what it flags.
If you decide to stop
"Uninstall openclaw" is searched more than most questions about it, so here is the short version from the uninstall guide. Preview first with openclaw uninstall --dry-run --all, then run openclaw uninstall --all. The interactive version only preselects the Gateway service, so pick state, workspace and app too if you want everything gone. Before you delete anything, move any workspace you want to keep out of ~/.openclaw, because removing state removes what is inside it. Then revoke the model keys and chat tokens it held at the provider's end, since deleting the files does not cancel them.
When a hosted AI employee fits better
OpenClaw's own trust model draws the line for you. If the people sharing the agent all trust each other, and someone technical is happy to own it, OpenClaw is a strong choice and it costs nothing to try. If the honest answer to "who will look after it" is "nobody really", the risks above are still there and nobody is watching them.
That is the gap a hosted AI employee is built for. With Sanaf, you pay for somebody else to be responsible for the parts above.
- Sign-ins are scoped per job and kept in an encrypted vault outside the model's reach, never in a prompt, a log or a config file on someone's laptop. You can revoke one without touching the rest.
- Approvals are enforced by the product. For each kind of action you choose whether it goes ahead, asks first or never happens, and it asks before anything it cannot undo.
- Tool output and email bodies are treated as content to work on, never as instructions to follow, and the consequential actions wait for a person anyway.
- There is one work log the team can read, so "what did it do while I was out" has an answer even when the person who set it up is on leave.
Sanaf works in Slack, Microsoft Teams and its own chat in the console. You give it jobs from a roster of thirty five, each with a written description: answering the support inbox, following up leads, booking appointments, chasing overdue invoices and the rest. You start with $120 of free credit and no card, then buy credits from $20 when you want them. They never expire and you pay for the work, not per seat. The pricing page has the details and the security page says how the vault and approvals work.
What you give up is real. Sanaf is not open source and you cannot self-host it. Your data is with a hosted provider, not only on your own machines. You can't swap in a local model or edit the runtime. If any of that is a hard requirement, OpenClaw is the honest answer, and our Sanaf vs OpenClaw comparison says so too.
If you are weighing other options, OpenClaw alternatives for teams covers four hosted products, and the best OpenClaw alternatives splits the list by whether you still want to run the agent yourself. Our post on open source AI agents and when hiring beats running one goes wider than OpenClaw.
Questions people ask about OpenClaw
Is OpenClaw safe to use?
For one person or a team that trusts each other, run on a dedicated machine with sandboxing on and a strong model, OpenClaw's defaults are careful and its docs explain every risk. The risks are what you would expect from an agent that can act: it uses whatever access you give it, any email or web page it reads can carry a prompt injection, and everyone who can message a shared agent shares its access.
Is OpenClaw free?
Yes. The software is MIT-licensed and the Foundation has no paid tier, no hosted service and no token. You pay your model provider for usage, pay for the machine it runs on, and spend your own time keeping it secure and up to date.
Who owns OpenClaw?
The OpenClaw Foundation, an independent 501(c)(3), stewards the project and employs the core team. OpenAI is one of several donors and employs its creator, but the FAQ says donors "do not own, control, or direct the project."
Can a team share one OpenClaw?
Yes. The docs call a shared gateway "a supported, first-class deployment", with shared sessions, per-person sign-in, named roles and a Slack or Teams bot. The limit is that one gateway is one trust boundary, so people who do not trust each other, or separate clients, need separate gateways.
Does OpenClaw work with Microsoft Teams?
Yes. Microsoft Teams is on its list of supported channels, alongside Slack, Discord, Google Chat, Telegram, WhatsApp and others.
What is the safest way to run OpenClaw at work?
Follow the checklist above: a dedicated machine with its own OS user, no personal accounts signed in, access through Tailscale or an identity-aware proxy, allowlisted channels with mentions required, sandboxing for non-admins, risky tools off by default, the strongest model you can afford, and openclaw security audit after every change.
What should a team use instead of OpenClaw?
If you like the idea but nobody wants to run it, a hosted AI employee like Sanaf in Slack or Microsoft Teams takes the hosting, the credentials and the approval rules off you. If self-hosting matters most, the best OpenClaw alternatives lists open-source agents you can still run yourself.