semantic-model-router

Smart LLM Router — routes every query to the cheapest capable model. Supports 17 models across Anthropic, OpenAI, Google, DeepSeek & xAI (Grok). Uses a pre-t...

This skill appears to be a local classifier that selects which provider/model would be most cost-efficient — it does not itself call provider APIs or require provider API keys. However: - SKILL.md's “zero external calls” is misleading: if sentence-transformers is installed and used, the encoder (all-MiniLM-L6-v2) will normally be downloaded from the internet (Hugging Face) unless already cached. If you require strictly offline behavior, do not install or instantiate the encoder, or ensure the model is pre-cached in a controlled environment. - The skill writes a local file named query_history.json by default; review/relocate it if you want to avoid storing queries on disk. - If you expected the skill to automatically route and invoke remote LLMs for you, note that it only selects/labels models and reports costs — you still need the provider integrations and credentials to actually call chosen models. Recommendations before installing: 1. Review the full model_router.py and model_weights.py contents locally (they are auditable). 2. If you want zero network activity, avoid installing/using sentence-transformers or pre-cache the embedding model in a private environment. 3. Run the skill in a sandbox or isolated environment first to observe whether it attempts to download models. 4. Confirm you understand that cost savings are estimates based on the embedded price catalog and that no provider billing or credential handling is included.

Type: OpenClaw Skill Name: semantic-model-router Version: 1.0.3 The skill bundle is classified as benign. The `SKILL.md` documentation accurately describes the skill's functionality and does not contain any prompt injection attempts. The `model_router.py` script implements the described LLM routing logic, utilizing local ML models and keyword matching. It performs file I/O to a `query_history.json` file for local logging, which is a plausible feature for 'refinement' and does not target sensitive paths. While the `sentence-transformers` library might perform a one-time model download on first use, this is a standard library setup behavior and not an 'external call' for classification inference, aligning with the stated intent of local operation and providing a fallback via `model_weights.py`.