Can you use OpenClaw without an API key? Every path, compared

OpenClaw itself never asks for money. Every reply it writes is a model call, though, and the API key is where that bill goes. So “without an API key” is really two questions: can the tokens be free, and can someone other than you hold the key. Both have answers. A fresh install makes the question concrete the first time you send a message with nothing configured:

No API key found for provider "openrouter".
Auth store: ~/.openclaw/agents/main/agent/openclaw-agent.sqlite
Configure an API key (openclaw models auth paste-api-key --provider openrouter)
| missing-provider-auth

Here is every way out of that error, with what each one costs and where it stops. Commands are from OpenClaw 2026.9.8; the upstream docs and every vendor page were read on 19 October 2026, and the links go to those pages.

The short answer

Path	Key to paste	Tokens cost	Runs on	The catch
A local model (Ollama, LM Studio, llama.cpp)	none	$0	your machine	the smallest managed llama.cpp recipe has an 8 GiB memory floor; small models truncate context and raise prompt-injection risk
Your ChatGPT subscription (Codex OAuth)	none, you sign in	your Plus or Pro plan	anywhere	one allowance shared with ChatGPT itself; eligibility is OpenAI’s call
Your Claude subscription (Claude CLI or setup-token)	none, a token from Claude Code	your Pro or Max plan	the host where Claude Code is logged in, or any host with a setup-token	Anthropic’s terms call API keys “the preferred way” for third-party tools and reserve the right to bill usage credits instead
Your GitHub Copilot plan	none, device login	your Copilot plan	anywhere	which models you see depends on the plan and on org policy
A free key (OpenRouter :free, Gemini API free tier, promos)	yes, a free one	$0	anywhere	50 requests a day on OpenRouter until you have bought $10; Gemini’s free tier uses your content “to improve our products”; promos expire
A host that holds the key (Aplite, Runware, Hostinger, Ampere, Everpod…)	none	credits or per token, billed by the host	their server	you pick from the host’s catalog, and a starter credit runs out
Bring your own key (self-hosting, xCloud, OneClickClaw, Flux…)	yes	the provider’s price	anywhere	this is the path the question is trying to avoid

1. A local model: no key, no token bill

The only path where nothing is billed per token. Ollama is the easiest runtime; LM Studio, llama.cpp and vLLM work through the same local models page.

1. Install Ollama from ollama.com/download and pull a model:

   ollama pull gemma4
   

2. Run onboarding and pick Ollama, then Local only:

   openclaw onboard
   

   Or set the model later:

   openclaw models set ollama/gemma4
   

3. The key field is a formality for local hosts. The Ollama setup page says loopback and private-network URLs “do not need a real bearer token” and that “any value works”:

   export OLLAMA_API_KEY="ollama-local"
   

Pick Ollama with the server not running and onboarding stops with Ollama could not be reached at http://127.0.0.1:11434. The limit is hardware and model quality: the upstream FAQ answers “Is a local model okay for casual chats?” with “Usually not”, because small or quantized models truncate context and raise prompt-injection risk. A 1 GB cloud machine is out; a desktop with 16 GB or more is in.

2. Your ChatGPT subscription

OpenClaw signs in to a ChatGPT account and spends that plan’s allowance instead of a key:

openclaw models auth login --provider openai

Add --device-code on a headless server. OpenClaw’s docs say OpenAI “explicitly allows subscription OAuth usage” in tools like OpenClaw, and OpenAI’s own developer page says “Eligible ChatGPT Plus and Pro users can use their ChatGPT plan for AI requests in participating apps” and that “ChatGPT plan usage is available to all open-source partners”. The catch is that it is one allowance: what the agent spends, ChatGPT no longer has, and the upstream FAQ notes that “Codex OAuth uses plan-dependent quota windows that can differ from the ChatGPT website or app”.

3. Your Claude subscription

Two token paths, both coming from Claude Code rather than from a key:

1. Claude CLI reuse. Claude Code is logged in on the same host; openclaw onboard offers it as a choice. OpenClaw prefers this path when it is available.

2. Setup token. On any machine with Claude Code:

   claude setup-token
   

   prints a long-lived token starting with sk-ant-oat01-, then:

   openclaw models auth login --provider anthropic --method setup-token
   

The policy is the part to read before relying on it. OpenClaw’s OAuth page says “Anthropic staff told us this usage is allowed again” and treats it as sanctioned “unless Anthropic publishes a new policy”. Anthropic’s own help article says API keys are “the preferred way to access Anthropic services using third-party software, tools, or services”, that Anthropic “may at its discretion allow paid subscribers who have enabled usage credits to use certain third-party tools”, and that it “reserves the right to draw use of such third-party tools from usage credits rather than subscription limits”. The Agent SDK article, updated 7 October 2026, says “You can still use the Claude Agent SDK, claude -p, and third-party apps with your subscription limits.” It works today for a personal agent; the discretion is Anthropic’s, and OpenClaw’s advice is: “For shared production automation, use an Anthropic API key instead of Claude CLI.”

4. Your GitHub Copilot plan

openclaw models auth login-github-copilot

runs GitHub’s device flow; the same login also serves embeddings for memory search. The Copilot page says “Copilot model availability depends on your GitHub plan and organization policy”, so what you can pick is whatever GitHub’s model picker shows your account.

There is no Google equivalent: the FAQ says OpenClaw “does not offer new Gemini CLI OAuth or Antigravity OAuth setup”, so Gemini means an AI Studio key. The Qwen, MiniMax and Z.AI coding plans are flat-price subscriptions that still hand you a key to paste.

5. A free key is still a key

Three ways to get one for $0, each with a ceiling:

• OpenRouter :free models. Paste a key from openrouter.ai; the free variants are limited to 20 requests a minute and 50 a day until you have bought $10 of credits, then 1,000 a day. One agent turn with tools is several requests, so 50 a day is a short afternoon.
• Gemini API free tier. Flash models are listed “Free of charge” on the pricing page, with the free tier marked “Content used to improve our products” and the paid tier “not used”.
• Promos. openclaw promos list shows time-limited free model offers published on ClawHub; openclaw promos claim <slug> sets the provider up and prints a sign-up URL for a free key. When the window ends “the provider stops serving the free models”.

6. A host that holds the key

Managed hosts that run models through their own account, so there is nothing to paste. Prices as each page read on 19 October 2026:

• Aplite. Models from the catalog over one metered account, billed in credits, 1 credit = 1¢; bringing your own key is not supported. The Free plan is one agent, budget models and 300 welcome credits, no card.
• Runware. Hosting free, inference “billed per token”; $2 in credit with a business email, no payment to deploy.
• Hostinger. “Pre-configured AI credits” in the subscription, amount not stated on the page, topped up through the AI Router from $5.99 for 5 credits; or “Use your ChatGPT Plus or Pro subscription”; from $5.99/mo on a 24-month term.
• Ampere. The FAQ: free hosting during beta, “$10 in AI credits included”, no card; the plan table lists “Free” at $1/mo with 5,000 credits at “$1 = 1,000 credits”. The two disagree.
• Everpod. $29/mo covering “up to $10 of model usage a month”.
• RapidClaw. 5 free messages, then $29/mo with $20 in API credits.
• Agent37. Free plan with “$0.20 of the default model to start”, then your own keys or a paid plan; Plus ($9.99/mo) bundles free models.
• OneClickClaw. No credits: bring a key, or sign in with a ChatGPT or SuperGrok subscription.
• KiloClaw. kilo.ai’s KiloClaw pages redirect to the homepage and the pricing page does not mention it, so there is nothing to check.

On Aplite the whole path is three steps: create an account with email or Google, press Deploy, ask. The usage shows as tokens under each reply and as money in Settings → Cost. What a managed instance gives up, the shell, raw openclaw.json and local models, and how the Free plan sleeps between uses, is on the OpenClaw page; the $0 self-hosted path is in OpenClaw hosting for free. Aplite is an independent service and is not affiliated with the OpenClaw project.

FAQ

Can I run OpenClaw without any API key at all?

Yes, with a local model: install Ollama, ollama pull gemma4, pick Ollama in openclaw onboard. Nothing is billed per token and no key is pasted; the cost is the machine, and small models answer worse than hosted ones. Every other path either signs in with a subscription or lets a host hold the key.

Can I use OpenClaw with my ChatGPT subscription?

Yes. openclaw models auth login --provider openai signs in with ChatGPT and spends the plan’s allowance; OpenAI’s developer docs say eligible Plus and Pro users can use their plan in participating apps, and open-source clients are included. The allowance is shared with ChatGPT itself.

Can I use OpenClaw with Claude Pro or Max?

It works through Claude CLI reuse or a claude setup-token, and OpenClaw’s docs say Anthropic has told them the usage is allowed. Anthropic’s help center calls API keys the preferred way for third-party tools and reserves the right to bill such use from usage credits rather than plan limits. For anything unattended, OpenClaw recommends an API key.

Why does OpenClaw say “No API key found for provider”?

The default model points at a provider with no auth profile. Paste a key with openclaw models auth paste-api-key --provider <id>, sign in with openclaw models auth login --provider <id>, or switch to a model whose provider is set up, for example openclaw models set ollama/gemma4.

Which OpenClaw hosts do not need my API key?

Hosts that meter models through their own account: Aplite, Runware, Hostinger, Ampere, Everpod, RapidClaw, and Agent37 from its Plus plan up. xCloud, OneClickClaw, Flux, ClawdHost and ClawHosters expect your own key, though OneClickClaw and Hostinger also take a ChatGPT subscription sign-in.