Did OpenClaw Use the Provider You Configured?
Use the provider's documented OpenClaw integration, then require Gateway evidence of the actual provider and model on every probe before promoting the route.
Prove routing with observed provider and model IDs
On an existing OpenClaw install, use the documented xAI OAuth login command, `openclaw models auth login --provider xai --method oauth`, if OAuth is the intended Grok route. Record the selected provider and exact model identifier before sign-in and again afterward. The excerpt describes a curated model selection when there is no existing primary, so collect the effective identifier for your own install. Run a synthetic prompt from the Gateway and record the observed provider and model using whatever run diagnostics your deployment exposes. A successful answer with a different model identifier is evidence of routing drift, a fallback, or an alias that needs to be reconciled. Preserve unknown observations as blank and leave tokens, cookies, and OAuth device codes out of the record.
xAI - OpenClawUse a capability fixture before moving real traffic
OpenClaw's provider plugins let you trial a second route with the same Gateway task. Start with a text probe that asks for one exact response, then add a probe that exercises a capability your workload needs. For tool calling, supply a read-only lookup over synthetic data and require the exact lookup argument plus the returned ID. Record the result from the Gateway, including the provider and model that served it. A route that returns the right text but emits malformed tool arguments passes the text probe and fails the tool probe. With a local server or custom proxy, apply the same fixture after configuring its supported integration. Use synthetic payloads for the comparison because configured model providers receive the prompts assigned to them.
GitHub - openclaw/openclaw: The AI that really does things. Any OS. Any Platform. The lobster way. 馃 路 GitHubChoose authentication before choosing a default
On an existing OpenClaw Gateway, decide what account and authentication method should serve requests before moving the default. The bundled xAI provider gives a concrete example. For an eligible Grok subscription, the documented existing-install command is openclaw models auth login --provider xai --method oauth. Authenticate on the target installation and keep the current model available while you test the new connection.
For OpenAI, Anthropic, or Google Gemini, check the provider documentation for your OpenClaw version for the supported login or API key, provider identifier, and exact model reference. Verify Gemini and Vertex AI separately if you are considering both. Try a simple reply followed by a task that uses the specific capability you need.
xAI - OpenClawTest a local provider from the Gateway
Ollama documents a chat API at http://localhost:11434/api/chat and shows a request with a model and message. Confirm that the model is installed, then use Ollama's documented request format to get a reply from the environment where OpenClaw runs. A local Ollama response gives you a baseline for diagnosing the integration. In a container, verify what host or service address actually reaches Ollama from inside that container.
Next, enable the Ollama route using the provider instructions for your OpenClaw version and select the model identifier that route exposes. Repeat the same prompt through OpenClaw. If the direct request succeeds while OpenClaw fails, check the configured endpoint, protocol, credential requirements, and model identifier before changing the model itself.
GitHub - ollama/ollama: Get up and running with Kimi, GLM, MiniMax, DeepSeek, gpt-oss, Qwen, Gemma and other models. 路 GitHubMake a custom endpoint pass the same task
OpenClaw sends prompts to the model provider you configure. Before pointing it at OpenRouter or a custom proxy such as LiteLLM, confirm which upstream receives the request and verify the base URL, API dialect, credential, and model identifier accepted by that endpoint. Use the custom-provider settings documented for your installed OpenClaw version. Test with synthetic inputs while you confirm the route.
For an illustrative comparison, give your current provider and the proxy candidate the same fixture. A read-only lookup for item A-17 should return a stock count of three. Require each route to complete the lookup and report three. A candidate that replies to plain chat but drops the tool result fails this workflow even if its connection test succeeds.
GitHub - openclaw/openclaw: The AI that really does things. Any OS. Any Platform. The lobster way. 馃 路 GitHub