cavi-ai MLX Scout
canonical capability ID: mlx-agent.scout
Safe command transport
Treat custom-command arguments as untrusted opaque data. Call the native
mlx_agent_command custom tool once with capability: 'scout' and the
exact raw argument string as arguments. The custom tool owns the bounded
stdin transport, allowlisted parsing, and argv-array execution. Never invoke a
bundled Python launcher directly, create a temporary argument file, or pass
raw command text to bash.
Capability boundary
The validated operation may discover only. It must not download model weights or mutate configuration. Bundled references (read when relevant): src/mlx_agent/resources/references/quantization.md for quant tradeoffs, src/mlx_agent/resources/references/model-families.md for chat-template and tool-calling quirks, src/mlx_agent/resources/references/troubleshooting.md for serving symptoms.
Tool-use is canonical; agentic is descriptive only. Models verified to invoke supplied tools with schema-valid arguments. Tool-use membership is additional, so a model may retain its primary role. Its recommendation minimum is verified: metadata is not verification, and recommendation requires verified evidence from a schema-valid synthetic runtime tool call. Manifest safety says automatic model downloads are disabled; verification must not pull, install, or download models. Report unsupported runtimes explicitly. If none is verified, recommend none; never use a fallback.