How to Connect Claude to OpenClaw

API key or Claude CLI: which should you use?

Both routes give OpenClaw the same Claude models. The difference is how you pay and where OpenClaw can run.

Set up Claude with an Anthropic API key

export ANTHROPIC_API_KEY="sk-ant-..."
openclaw onboard --anthropic-api-key "$ANTHROPIC_API_KEY"

openclaw models list --provider anthropic
openclaw models set anthropic/claude-opus-5-5
openclaw models status --probe

Set up Claude with your Claude Code sign-in

OpenClaw runs the installed Claude Code program for each turn. It never reads, stores or refreshes your Claude login tokens. Claude Code keeps control of them.

claude --version
claude auth status --text
claude auth login     # only if you're not signed in
claude update         # if OpenClaw says the build is incompatible

openclaw onboard
# choose: Claude CLI

{
  agents: {
    defaults: {
      model: { primary: "anthropic/claude-opus-5-5" },
      models: {
        "anthropic/claude-opus-5-5": {
          agentRuntime: { id: "claude-cli" },
        },
      },
    },
  },
}

Use a Claude setup token

On any machine with Claude Code, claude setup-token prints a long-lived token starting with sk-ant-oat01-. Store it in OpenClaw:

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

In the macOS app, choose Anthropic setup-token under Connect with an API key or token. Setup tokens are still supported, but OpenClaw prefers Claude CLI reuse when it's available. Tokens can expire or be revoked, so for new setups the docs suggest an API key.

Choose a Claude model

Use the full reference, or one of the short aliases, with openclaw models set. Fresh setups default to Opus 5.5.

*US dollars per million tokens for API-key billing, as listed in OpenClaw's model catalog. Check Anthropic's pricing page for current rates. All five models have a 1,000,000-token context window and up to 128,000 output tokens.

Claude config builder

Control how hard Claude thinks

Current Claude models use adaptive thinking: they decide how much to reason, guided by an effort level. Change it from any chat:

/think low       # faster and cheaper
/think high      # harder problems
/think default   # back to the model's default

• Sonnet 5.5 also accepts /think off, which turns off up-front thinking.
• Opus 5.5 and Fable can't turn thinking off. A saved off setting becomes low.
• Changing the model or thinking level mid-session can drop Claude's earlier reasoning. Pick both when you start a session if continuity matters.

Prompt caching and cost tracking

With an API key, OpenClaw caches the repeated part of each prompt automatically. Cache reads cost a fraction of normal input.

{
  agents: {
    defaults: {
      models: {
        "anthropic/claude-opus-5-5": {
          params: { cacheRetention: "long" },
        },
      },
    },
  },
}

Fast mode, compaction and refusal fallback

/fast on
/fast off

{
  agents: {
    defaults: {
      models: {
        "anthropic/claude-sonnet-5-5": {
          params: { anthropicServerCompaction: true },
        },
      },
    },
  },
}

openclaw models fallbacks add <provider/model>
openclaw models fallbacks list

Fix common Claude connection errors

More fixes: troubleshooting guide.

Claude and OpenClaw questions