OpenClaw pricing: free software, but what do you actually pay?

Is OpenClaw actually free?

Yes, for the software itself. OpenClaw ships under the MIT license, and our tool page lists a single plan, called Open source, at $0 per month. The official site and the GitHub repository both state that there is no paid tier, no hosted service and no token. So there is no Pro version waiting to be unlocked, no seat to buy, no credit tier to cross once usage picks up. That is why its free offering earns the top mark of 5 out of 5 in our review. The project is run by the OpenClaw Foundation, a US 501(c)(3) nonprofit, and the code is public at github.com/openclaw/openclaw. Everything we verified sits on the OpenClaw tool page.

So where does the real cost come from?

From two line items the publisher does not bill, but somebody does: the AI model and the machine. OpenClaw runs with the model you pick, hosted by a provider or executed locally. In the first case you plug in your own API key, and that provider bills you directly for what you consume. The agent also has to run somewhere: your own computer, or a server you rent.

We score the price criterion 4 out of 5: the software is free, but AI models are billed separately.

Can you put a number on the bill before you install?

Not from published data. OpenClaw cites no hosting price and no model provider price, and we carry none either: those amounts are not verified in our database. Three variables drive the spend. The model first, hosted or local. The message volume second: the agent is driven from 20+ messaging apps (WhatsApp, Telegram, Slack, Discord, iMessage, Signal), which makes it very easy to reach and mechanically raises the number of exchanges it processes. The machine last, depending on whether you reuse an existing computer or rent a server.

In our reading, the second variable is the one people underestimate. An agent you can ping from WhatsApp gets used far more often than a tool that requires opening an interface, and it is the model provider’s meter that records the difference. Before you open access to a whole team, set a spending cap with your model provider.

Where is your data stored?

With you, for the most part. State, memory and credentials stay on the machine running OpenClaw, in the ~/.openclaw folder. Messages processed by the AI do leave for the provider of the model you picked, unless you run a local model. That design is why we score the privacy and GDPR side 4 out of 5: where the data sits depends entirely on you, since the publisher runs no hosted service.

The logical consequence is that no GDPR policy and no DPA are published, and we will not invent their terms. On what goes back to the foundation, a daily version check is on by default (version, operating system, Node.js, architecture) and remains possible to turn off; anonymous statistics are opt-in.

What are the risks of an agent that acts on your machine?

The caveats are documented and serious. By default, the agent’s tools run directly on the host machine, with no sandbox. And because the agent reads incoming messages, those messages can carry prompt injection: security is yours to configure, setting by setting. On the credit side, a security audit was carried out by Trail of Bits, and the documentation publishes a security section plus an explicit trust model.

Put another way, free software does not remove the scoping work. Running the agent on a dedicated machine, limiting the access you hand it and reading the trust model before go-live are part of the entry cost, even if that cost never lands on an invoice.

What do our scores say, criterion by criterion?

The gap between 5 out of 5 on features and 2 out of 5 on ease of use sums up the case: the technical capability is there, and so is the implementation load.

Who is OpenClaw a good fit for?

A technical profile: someone hosting a personal agent who takes on its permissions, its models and its security. If you already administer a server, if you know how to create and rotate API keys, and if you are comfortable deciding yourself what a program acting on your system may touch, the ratio between what the tool can do and what it costs is favorable.

For a small business owner with no technical resource in house, the answer is different. A 2 out of 5 on ease of use is not a detail: nobody will answer in your place on the absence of a sandbox by default. If your need is about automating business processes rather than running a personal assistant, look at automation tools instead, n8n and its free self-hostable Community Edition among them. Our guide to self hosting n8n covers what that involves, and the one on how to build an AI agent starts from the need rather than from the tool.

In between sits a defensible middle case: a technical freelancer who wants an assistant reachable from Telegram or Slack, with memory and credentials on their own machine. That is the ground where OpenClaw does what it says, provided you keep an eye on the model provider’s meter.