OpenClaw Model Auto Router

Problem

When OpenClaw’s /model output contains multiple models, the issue is not usually “no model available,” but choosing the right one for the current task: Chinese code, English reasoning, long-form Q&A, vectorization, or low-cost quick answers may require different models. Manually comparing options is slow and can mix chat models with embedding models.

How It Works

model-auto-router turns routing into concrete steps:

• Get the model list: Prefer the /model result already present in the conversation context; if absent, ask the user to provide it.
• Two-level classification: Use keyword rules first for millisecond-level matching. For unmatched unfamiliar models, run parallel web_search requests and extract parameter scale, strong tasks, and official positioning. Fall back to BALANCED when information is insufficient.
• Analyze the task: Consider language, complexity, task type, and cost sensitivity, including code/math/reasoning, creative writing, Q&A, vectorization, and general tasks.
• Pick within the bucket: Prefer non-:free models when cost is not a concern; prefer :free models when cost-sensitive. Within the same bucket, prefer larger parameters or newer versions when conditions are equal.
• Output the result: Provide the recommended model, routing reason, and switch command. If multiple candidates exist, include alternatives. If the target bucket is empty, fall back to a default bucket and explain.

Boundaries

The skill does not introduce models on its own and does not rely on external APIs; its core input is the model list recognized by OpenClaw /model. The classification table is cached within the session and invalidated when the user explicitly says the list changed or a new /model run shows a different list. If the user specifies a model, routing is skipped. Embedding models are reserved for embedding requests and do not participate in ordinary chat routing.