How to Configure OpenClaw Anthropic

Anthropic configuration in OpenClaw is straightforward if you separate auth from runtime. The docs support direct Anthropic API keys and Claude CLI credential reuse. For long-lived Gateway hosts, API keys are the clearest production path. Claude CLI reuse is useful when you already have a local CLI login and intentionally want OpenClaw to route through that backend.

30-second answer

For API-key auth, create an Anthropic Console key, run onboarding or pass --anthropic-api-key, and set a model such as anthropic/claude-opus-4-6. For Claude CLI reuse, install Claude CLI, confirm claude --version, run onboarding, and use the canonical anthropic/* model ref with agentRuntime.id: claude-cli when you want that execution backend. Legacy claude-cli/* refs remain compatibility paths, not the preferred new config.

Where it fits

Use Anthropic when Claude is the right reasoning model for your agent workload, especially long-form planning, review, and careful tool use. Use API keys when billing and production control matter. Use CLI reuse for local operator workflows where the host already has Claude CLI auth and you accept the operational dependency on that local login.

Docs-grounded facts

• OpenClaw supports Anthropic API-key auth and Claude CLI reuse.
• API keys use anthropic/* model refs.
• Claude CLI reuse should use canonical anthropic/* refs plus agentRuntime.id: claude-cli for new config.
• Claude 4.6 defaults to adaptive thinking when unset.
• Prompt caching supports short, long, and none for API-key auth.
• The docs call API keys the clearest production path for long-lived gateways.

Set it up deliberately

The docs show config with env ANTHROPIC_API_KEY and agents.defaults.model.primary for API-key auth. For CLI reuse, the preferred shape keeps provider/model selection as anthropic/* and places execution backend choice in agentRuntime.id. Anthropic also supports prompt caching for API-key auth, with cacheRetention short, long, or none configured under model params.

Use it safely

Do not confuse Claude CLI auth with Anthropic API-key auth. They are different routes with different operational failure modes. For Claude 4.6, OpenClaw defaults thinking to adaptive when no explicit level is set, but you can override per message with /think or in model params. Cache settings should match workload privacy and reuse expectations.

Common mistakes

The common mistake is using legacy claude-cli/* model refs in new config instead of canonical anthropic/* plus agentRuntime.id. Another mistake is assuming CLI reuse is the best production choice because it is convenient locally. The docs explicitly call API keys the clearest and most predictable path for long-lived Gateway hosts.

Verification checklist

Run openclaw models list --provider anthropic, send a small test, and confirm the route. If using Claude CLI, verify the CLI is installed and logged in from the same host context the Gateway uses. If prompt caching matters, test with model params and watch provider behavior or usage reporting for the expected cache retention.

Playbook angle

The OpenClaw Playbook’s Anthropic advice is boring on purpose: pick API key or CLI, write it down, set model refs canonically, and avoid mixing provider names with runtime names. That keeps future agent debugging sane.

Operator note

How to Configure OpenClaw Anthropic works best when it is written into a small runbook instead of left as tribal knowledge. Record the intended owner, the exact config surface, the channel where results should appear, the allowed inputs, the expected output, and the rollback step. OpenClaw gives agents broad tools, but the durable value comes from making each tool boring, repeatable, and auditable. I would rather have one well-scoped Anthropic config workflow that survives a restart than five clever demos nobody can safely run next week. If the runbook cannot explain when not to use it, keep refining before automation becomes default.