Do you need an API key to use OpenClaw?
Yes: OpenClaw brings the harness, you bring the model. What a model key is, where to get one, and what it costs.
Mostly yes, with one big 2026 exception. OpenClaw is the harness, the body of the agent, but the thinking is done by an AI model, and you connect one yourself. The classic way is an API key from a model provider (Anthropic, OpenAI, and many others), billed by usage. Since May 2026 there’s a keyless alternative: signing in with a ChatGPT subscription, which powers the agent from the plan you already pay for.
Why OpenClaw doesn’t come with a model
OpenClaw is open-source software that you run yourself. There’s no OpenClaw company account, no subscription, and no built-in model, which is a feature, not a gap. It means your agent isn’t tied to one AI vendor: you choose the model, you can change it later, and your conversations flow between you, your machine, and the provider you picked. This is usually called bring your own key, or BYOK.
What an API key actually is
An API key is a long secret string (it looks something like sk-ant-api03-…) that acts as a combined username, password, and billing account for a model provider. When your agent wants to think, OpenClaw sends the request to the provider along with your key, and the provider bills your account for the tokens used. Two things follow from that:
- Treat it like a bank card number. Anyone who has your key can spend your money. Don’t paste it into random websites, don’t commit it to Git, and set a spending limit in the provider’s dashboard. Every major provider has one.
- Costs are usage-based. There’s no flat fee: light personal use of a mid-size model typically costs a few dollars a month, while heavy agent workloads on frontier models can cost much more. The spending limit turns “can cost much more” into a number you chose.
Which providers work?
OpenClaw supports the major model providers: Anthropic (Claude), OpenAI, Google (Gemini), Mistral, and a long list of others, plus router services like OpenRouter that put many models behind one key, and local runners like Ollama for models on your own hardware. You pick a provider during openclaw onboard, and your default model is set as provider/model in the config, so switching later is a config change rather than a migration.
If you’re unsure where to start: pick the provider whose models you already know, create the key, set a monthly cap, and move on. The choice isn’t permanent, and agonizing over it is the least productive hour of any OpenClaw setup.
The keyless path: subscription sign-in
The exception worth knowing before you fetch any key: OpenClaw can sign in with your ChatGPT account, officially supported by OpenAI since May 2026 and available on every plan including Free. No key, no separate billing: the agent draws on your plan’s included Codex usage. Two caveats: subscription allowances come in rolling windows (heavy agent days can exhaust one), and a few subsystems (memory embeddings, voice) still want an API key or an alternative provider. Claude subscriptions are the cautionary opposite: that story is complicated, and an API key remains the predictable way to use Claude models.
The clunkiest step, and the ways around it
Fair warning: getting the key is the clunkiest step of setting up a self-hosted agent: a tour through a provider dashboard, a billing page, and a copy-paste, usually easiest from a real computer. It’s a one-time errand, and for self-hosters it’s simply part of the deal. It’s also a step managed hosting can remove outright: an Everpod pod comes with up to $10 of model usage included every month, so there’s no key to fetch before you meet your agent. If you prefer running on your own key (your provider account, your usage bill), that stays fully supported, with nothing added on top.