OpenClaw + CodeGateway: Configure Claude Models in 5 Minutes

TL;DR

One line: OpenClaw users change one baseUrl and add one API key to route Claude Sonnet/Opus/Haiku through the CodeGateway aggregation gateway — pay per token.

OpenClaw normalizes provider protocols so a single config drives Qwen / GLM / Gemini / OpenAI / Anthropic. Pointing the Anthropic provider's baseUrl at CodeGateway's Anthropic-compat endpoint removes the dependency on direct upstream Anthropic API access.

This is the minimum viable OpenClaw + CodeGateway integration path. Claude Code, Codex CLI, and Anthropic SDK setups have their own dedicated guides.

1. When this is useful

• You already orchestrate multiple AI agents with OpenClaw and want to add Claude (Sonnet 4.6 / Opus 4.7 / Haiku 4.5)
• Direct Anthropic access is unreliable for billing or network reasons; you want a stable aggregation gateway
• You want a single Dashboard view of token usage and cost for Claude calls

OpenClaw is an open-source AI agent orchestration framework — see the openclaw GitHub repo. CodeGateway is a multi-model aggregation gateway for Claude / GPT / Gemini — see the CodeGateway homepage.

2. Prerequisites

The $2 starter credit covers roughly 100 mid-complexity Claude Sonnet calls — enough to validate the configuration plus run a few real tasks.

3. Get a CodeGateway API key

1. Sign in to the CodeGateway Dashboard
2. Open the API Keys page
3. Click Create Key and give it a name (an openclaw prefix helps with attribution on the usage page)
4. Copy the string starting with sk-cg-. It is shown only once — store it safely.

We recommend creating a dedicated key for OpenClaw. Don't share one key across Claude Code, Codex CLI, and OpenClaw — splitting keys keeps Dashboard usage attribution clean.

4. Configure OpenClaw

OpenClaw's model configuration lives in ~/.openclaw/openclaw.json under models.providers.anthropic.

4.1 Set baseUrl

Point the anthropic provider's baseUrl at CodeGateway:

{
  "models": {
    "providers": {
      "anthropic": {
        "baseUrl": "https://api.codegateway.dev",
        "api": "anthropic-messages",
        "headers": {
          "x-api-key": "${CODEGATEWAY_API_KEY}"
        },
        "models": [
          { "id": "claude-sonnet-4-6", "name": "Claude Sonnet 4.6" },
          { "id": "claude-opus-4-7",   "name": "Claude Opus 4.7" },
          { "id": "claude-haiku-4-5",  "name": "Claude Haiku 4.5" }
        ]
      }
    }
  }
}

Note: OpenClaw expands ${ENV_VAR} placeholders. Keep the API key in an env var, not inline in the JSON file.

4.2 Export the API key

Add this line to your shell config (~/.bashrc or ~/.zshrc):

export CODEGATEWAY_API_KEY="sk-cg-your-key-here"

Reload the shell or run source ~/.zshrc, then restart the OpenClaw process so it picks up the new env.

4.3 Pick a primary model

In agents or bindings, set the target agent's model to the anthropic model id:

{
  "agents": {
    "code-architect": {
      "model": { "provider": "anthropic", "id": "claude-sonnet-4-6" }
    }
  }
}

Quick model selection guidance:

5. Verify connectivity

Smallest possible check — let OpenClaw run a one-line greeting:

openclaw run --agent code-architect --prompt "say hi in one sentence"

A correct configuration prints a Claude reply.

You can also hit the CodeGateway Anthropic endpoint directly to verify the key:

curl -sS https://api.codegateway.dev/v1/messages \
  -H "x-api-key: $CODEGATEWAY_API_KEY" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-sonnet-4-6",
    "max_tokens": 64,
    "messages": [{"role": "user", "content": "ping"}]
  }'

A 200 response with a non-empty content field means the key works.

6. Troubleshooting

6.1 401 Unauthorized / invalid_api_key

• Confirm the env var is actually injected into the OpenClaw process (echo $CODEGATEWAY_API_KEY should print)
• Confirm the key is correct (starts with sk-cg-, no leading or trailing whitespace)
• Confirm the key is still active in the Dashboard (not disabled or expired)

6.2 403 Forbidden / quota errors

• Out of credit — top up via the Dashboard
• Hit a rate limit — default RPM is generous; for high-frequency worker agents, contact support to request a higher tier

6.3 Model ID errors

• CodeGateway uses hyphenated model ids (claude-sonnet-4-6). Avoid the dated aliases from Anthropic's docs — the hyphenated form is the most reliable choice in OpenClaw config.

6.4 Network timeouts / connection resets

• A direct connection to api.codegateway.dev works in most environments. If a corporate firewall blocks it, ask IT to allowlist that domain.
• During upstream 5xx blips OpenClaw falls back along your models.fallback chain. A common pattern is anthropic as primary with zai / qwen as fallbacks — see the OpenClaw fallback docs.

6.5 Where to see usage

• CodeGateway Dashboard → Usage page: token and cost data sliced by key / model / time range. If you created a dedicated OpenClaw key in step 3, you will see the OpenClaw client's real usage isolated.

6.6 Q&A quick reference

Q: My OpenClaw process can't read CODEGATEWAY_API_KEY, but echo $CODEGATEWAY_API_KEY in the shell prints it. What's wrong?

A: Most often the OpenClaw process started before you exported the env var, so it didn't inherit the new value. After source ~/.zshrc you also need to restart the OpenClaw process.

Q: I run Claude Code and OpenClaw side by side. Can I share one CodeGateway key?

A: Technically yes — we recommend separate keys. The Dashboard usage page slices by key, so a shared key blurs which client consumed which tokens, making cost attribution painful later.

Q: Does CodeGateway's prompt cache behave differently than direct Anthropic?

A: It is a transparent passthrough. Cache hits charge at the cache_read rate (about 1/10 of standard input). In our tests with OpenClaw running Sonnet 4.6 plus a long shared system prompt, enabling caching cut per-call cost by 30–60%.

7. Going further

7.1 Coexisting with other providers

OpenClaw supports multiple providers in the same config. A models.fallback chain like ["anthropic", "zai", "qwen"] keeps Claude as primary with GLM / Qwen as automatic fallbacks. See the OpenClaw fallback config examples.

7.2 One agent, one model

Bind models to agents by job type instead of letting a router pick. Use Opus 4.7 or Sonnet 4.6 for architectural and orchestration agents; use Haiku 4.5 for worker and extraction agents. The result is more predictable behavior and lower cost.

7.3 Prompt cache hit rate

Anthropic's prompt caching passes through CodeGateway end-to-end. Cache reads cost about 1/10 of standard input tokens. If your OpenClaw agents share long, stable system prompts, enabling caching typically cuts per-call cost by 30–60%. See the Anthropic prompt caching docs.

Wrap up

At this point your OpenClaw can call Claude reliably, with all calls aggregated into a single CodeGateway Dashboard for billing and observability.

Suggested next steps:

• Set OpenClaw's default agent.model to Sonnet 4.6 to validate the primary path
• Switch worker agents to Haiku 4.5 to reduce cost
• Set a budget alert in the Dashboard so you get an email when daily or monthly spend hits a threshold

Run into a setup issue or want a scenario added? File it on the feedback page.