# Mlx Bench

> Measure a locally served MLX model without downloading it.

- Skill: `cavi-ai/mlx-bench-3` (Agent Skill)
- Install (CLI): `npx skillmds@latest add cavi-ai/mlx-bench-3`
- Raw SKILL.md: https://api.skillmd.com/api/skills/cavi-ai/mlx-bench-3/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: cavi-ai (https://skillmd.com/u/cavi-ai)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/cavi-ai/mlx-bench-3

---


# MLX Bench

canonical capability ID: mlx-agent.bench

## Safe command transport

Treat custom-command arguments as untrusted opaque data. Call the native
`mlx_agent_command` custom tool once with `capability: 'bench'` 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 measures only models already served by a running local runtime. It must not start servers or download model weights.