400+ AI models
in every OpenClaw agent, pay per token.
Register one tool in OpenClaw and every model in the catalog — GPT-4, Claude, Gemini, Llama, and hundreds more — becomes a single parameter in your agent. No vendor accounts to manage, no API keys to rotate, no per-model billing to reconcile.
Real OpenClaw agents that need more than one model.
Each pattern below is an OpenClaw agent with the AI model gateway registered as one tool. The math works because each model call costs only the tokens it generates.
Send each step to the right model.
An OpenClaw agent that classifies intent doesn't need the same model as the one that writes the final output. Register the gateway once and let the agent pick the model per step — lightweight classification to a fast cheap model, complex drafting to a frontier one. One tool, one budget line, regardless of how many models the agent touches.
Compare model outputs inside the agent.
An OpenClaw evaluation agent sends the same prompt to GPT-4, Claude, and Gemini in parallel tool calls, then scores and selects the best response. One tool registration covers all three — changing the model parameter is the entire diff between branches. Per-token billing keeps comparison runs proportional to what they actually generate.
Pick the retrieval model separately from the generation model.
An OpenClaw RAG agent can use a fast model to retrieve context and a more capable model to generate the final answer — all through the same tool. The agent declares the model per step; OpenClaw budget caps keep total spend predictable across both retrieval and generation calls.
Run structured tool calls across any model.
An OpenClaw agent that needs tool calling or JSON mode can route those steps to models that support them — without registering a different tool for each. Pass tool definitions the same way you would for OpenAI; the gateway routes them and returns responses in the same OpenClaw-parseable format. Idle runs cost zero.
One tool. Every model your agent will ever need.
Work with familiar patterns, including: - Standard chat completions - System, user, and assistant message roles - Generation controls such as temperature and top_p - Function and tool calling (on supported models) - JSON mode for structured outputs
- Single OpenClaw tool
- 400+ models
- Budget caps honored
- OpenAI API compatible
OpenClaw-specific questions.
If something below doesn't cover your case, ping us — we work directly with OpenClaw builders, no SDR funnel.
How does this register as an OpenClaw tool?
+
It's a standard POST endpoint that accepts the OpenAI chat completions body. Register it in OpenClaw the same way you'd register any HTTP tool — endpoint, schema, per-token price. OpenClaw uses the price to enforce budget caps and to show the user what each model call will cost before the agent runs.
Does OpenClaw need separate API accounts for each model?
+
No. The gateway holds the vendor relationships. OpenClaw pays per token from a wallet you connect — no OpenAI account, no Anthropic account, no Google AI account to create or maintain. One key, one billing line, regardless of how many models your agents call.
How does the token ceiling interact with OpenClaw budget caps?
+
They stack. Each tool call sets a max_tokens ceiling so no single generation step can overspend. OpenClaw run budget caps then bound the total spend across all tool calls in an agent run. The agent stops issuing new calls when either limit is reached.
Can an OpenClaw agent call different models in the same run?
+
Yes — that's the primary design point. The model is a parameter on each tool call. A single agent can call GPT-4 for reasoning, Llama for classification, and Gemini for summarization without registering three tools or managing three API keys.
Does tool calling work in OpenClaw agents?
+
Yes, on models that support it. Pass tool definitions in the same format as OpenAI. The gateway routes them and returns tool call responses in the same structure OpenClaw already parses.
What happens when a model is deprecated?
+
Update one parameter in the agent — the model name. No tool re-registration, no new API account, no integration work. New models are added to the catalog automatically and become available to your agents without code changes.