OpenClaw with Claude: Configuration Guide

OpenClaw ships with Claude support built in. If you're running DenchClaw and want to use Anthropic's models — known for careful reasoning, strong writing quality, and reliable instruction following — this guide walks you through the complete setup.

Why Claude for DenchClaw?#

Claude's models have distinct characteristics that make them particularly well-suited for CRM and agent workflows:

Strong instruction following: Claude reliably does what you ask without adding unsolicited content or ignoring constraints. For agent loops where the model must follow a specific format or stop condition, this consistency matters.

Thoughtful output: Claude tends to produce well-structured, readable responses. For customer-facing outputs like outreach emails and proposals generated by Skills, Claude's prose quality is consistently high.

Extended thinking: Claude's latest models support explicit reasoning traces, letting you see why the model made a decision — useful for debugging complex Skills.

Safety-conscious by design: Claude is trained to avoid harmful outputs and flag uncertainty. For a CRM agent with access to customer data, having a model that errs on the side of caution is generally a feature, not a bug.

Long context: Claude handles 200K+ token context windows, enabling full-document processing and large dataset analysis.

Step 1: Get Your Anthropic API Key#

1. Go to console.anthropic.com
2. Sign in or create an account
3. Navigate to API Keys
4. Click Create Key
5. Copy the key immediately — you can't retrieve it later

Make sure your account has credits or an active subscription. Anthropic uses a prepaid credit system for API access.

Step 2: Add the API Key to OpenClaw#

openclaw config set apiKeys.anthropic YOUR_ANTHROPIC_API_KEY

Or in ~/.openclaw/config.json:

{
  "apiKeys": {
    "anthropic": "sk-ant-..."
  }
}

Step 3: Configure Claude as Your Model#

openclaw config set model anthropic/claude-sonnet-4-5

Or in the config:

{
  "model": {
    "provider": "anthropic",
    "model": "claude-sonnet-4-5"
  }
}

Restart to apply:

openclaw restart

Step 4: Verify the Setup#

openclaw chat "What contacts were added to my CRM this week?"

You should see a Claude-style response — thorough, well-formatted, and directly answering the question. If you get an API error, double-check the key format (should start with sk-ant-).

Choosing the Right Claude Model#

Anthropic maintains a tiered lineup with different speed/capability tradeoffs:

For DenchClaw daily use: claude-sonnet-4-5 is the sweet spot. It's fast enough for interactive queries and capable enough for complex Skills. Most users won't need to deviate from this.

For complex analysis: Switch to claude-opus-4 when you're asking the agent to do something genuinely hard — analyzing large datasets, synthesizing research, or reasoning through a multi-step problem that previous attempts got wrong.

For high-volume automation: Use claude-haiku-3-5 when you're running bulk operations (enriching hundreds of contacts, processing many records) where cost and speed matter more than reasoning depth.

Setting Per-Task Model Overrides#

DenchClaw lets you override the model per command:

# Quick lookup — use Haiku
openclaw chat --model anthropic/claude-haiku-3-5 "How many deals in my pipeline?"
 
# Strategic analysis — use Opus
openclaw chat --model anthropic/claude-opus-4 "Review my last 6 months of sales data and identify the key patterns"

This lets you default to a cost-efficient model and selectively upgrade when the task warrants it.

Enabling Extended Thinking#

Claude's latest models support an "extended thinking" mode where the model works through problems step by step before responding. This improves accuracy on complex reasoning tasks.

In OpenClaw, you can enable thinking-mode reasoning:

openclaw config set model.thinking true

Or toggle it per-session:

openclaw chat --thinking "Analyze my pipeline and tell me which deals are most likely to slip"

Extended thinking is slower and costs more tokens, so use it selectively — for analysis tasks where getting the right answer matters more than speed.

Claude and DenchClaw's Skills System#

Claude is particularly strong at the Skills system's core operations. Here's why:

Reliable tool calling: Claude follows function calling schemas precisely, which is critical for Skills that need to write data back to DuckDB or trigger external integrations.

Context-aware behavior: Claude maintains coherent context across long agent sessions. If you start a complex research task and it requires multiple steps, Claude tracks state well.

Safe defaults: Claude won't take destructive actions without confirmation. If a Skill might delete records or send emails, Claude will ask rather than assume.

For building custom DenchClaw Skills, Claude is a strong choice for the development loop — it follows the SKILL.md format instructions precisely.

Using Claude's Long Context for CRM#

Claude's 200K context window is genuinely useful for CRM workflows:

Full history analysis: Feed in an entire customer conversation history (emails, notes, call logs) and ask Claude to summarize the relationship and suggest next steps.

Document processing: Attach large contracts, proposals, or decks to a CRM contact and have Claude extract key information and populate fields automatically.

Bulk review: Include your entire lead list in context and ask Claude to score and prioritize based on criteria you specify.

This level of context capacity means you rarely need to worry about hitting limits for typical CRM data volumes.

Combining Claude with Local Data Storage#

Like all model integrations in DenchClaw, your CRM data in DuckDB stays on your machine. When you run a Claude query:

1. OpenClaw pulls relevant data from local DuckDB
2. Constructs a prompt with that data
3. Sends it to Anthropic's API
4. Stores the response back in your local DuckDB if configured

The data leaving your machine is only what you explicitly include in the query context. Your full database never gets sent wholesale.

If you need complete data privacy with no API calls at all, see the Ollama setup guide for a fully local configuration.

Cost Management#

Anthropic uses a token-based pricing model. For DenchClaw workloads:

• Interactive queries (short context, short response): fraction of a cent each
• Complex analysis (long context, detailed response): a few cents
• Bulk enrichment (many records): can add up — use Haiku or batch carefully

Set a budget in your Anthropic console and monitor usage under Usage in the sidebar. For high-volume workloads, build in cost checkpoints — pause and review if daily spend exceeds a threshold.

Troubleshooting#

"Overloaded" or 529 errors

Anthropic occasionally gets capacity-constrained on high-demand models. Retry after a few seconds. If it's persistent, switch to Haiku temporarily — it typically has more capacity headroom.

Responses are much longer than expected

Claude tends toward thorough responses. If your use case needs brevity, add an explicit instruction to your system prompt: "Be concise. Responses should be 2-3 sentences unless detail is requested."

Tool calling not working as expected

Verify you're using a model that supports tool use (Opus, Sonnet, and Haiku all do). Check that the tool schema is valid JSON. Claude is strict about schema conformance — malformed schemas cause silent failures.

Rate limit errors

Anthropic rate limits are per-API-key. If you're hitting limits during bulk operations, add delays between requests or request a rate limit increase from Anthropic support.

FAQ#

Is Claude better than GPT-5 for DenchClaw?

They have different strengths. Claude tends to be more consistent and follows instructions more precisely; GPT-5 has a broader ecosystem and more fine-tuned variants. For writing tasks and careful reasoning, many users prefer Claude. For code generation and tool-heavy Skills, GPT-5 is often stronger. Try both with your actual workload — the difference depends on your specific use case.

Can I use Claude through AWS Bedrock with OpenClaw?

Yes. AWS Bedrock provides Claude models with the same API format. Set your baseUrl to your Bedrock endpoint and configure AWS credentials. This is useful for teams already on AWS infrastructure with data residency requirements.

Does Claude work with all DenchClaw Skills?

Yes. All built-in Skills are tested against Claude Sonnet. Custom Skills should work if they use standard tool calling — just verify the schema format matches Anthropic's specification.

How does extended thinking affect cost?

Extended thinking uses more output tokens (for the thinking trace) and takes longer. For a task that normally costs 0.5 cents, thinking mode might cost 2-3x that. Worth it for genuinely hard problems, not for routine queries.

Can I use Claude for both the default model and specific Skills simultaneously?

Yes. You can configure claude-sonnet-4-5 as your default and set specific Skills to use claude-opus-4 for tasks that benefit from deeper reasoning. Skill-level model overrides are configured in each Skill's settings.

Ready to try DenchClaw? Install in one command: npx denchclaw. Full setup guide →