Talk to your coding agent. It talks back. Local-first voice I/O (MCP tools + Agent Skill) for Claude Code, Cursor, OpenClaw and any MCP client.
VoxAgent defines 3 tools with clear, action-oriented names (listen, speak, ask_user) and good descriptions (avg 180 chars, within 10-1024 baseline). All tools have descriptions and input schemas are present. However, input parameters lack type annotations in the schema (listen has empty {}, speak and ask_user have minimal schema). The 'listen' tool has no parameters but its schema is empty object rather than properly typed. Descriptions are well-written and contextual but lack explicit constraint documentation (e.g., text length limits for speak). No output schemas are documented. Error handling is minimal, tools return fallback strings like '[nothing heard]' but don't guide recovery. No tool annotations (readOnlyHint, destructiveHint) despite all being read-only operations.
Ask the user a question out loud and wait for their spoken answer. Returns the transcript of what the user said. Use for quick confirmations or choices while working hands-free, e.g. ask_user('Run the tests now?').
Record speech from the default microphone and return the transcript. Use this to hear the user: call it when the user asks to dictate, when a task needs verbal input, or as a hands-free way to receive the next instruction.
Say `text` out loud through the speakers. Returns seconds spoken. Keep spoken text short (1-2 sentences). Ideal for progress updates, asking a quick question, or reading a summary aloud while the user is away from the keyboard.
Input schemas lack proper type definitions. 'listen' has empty schema {}; 'speak' and 'ask_user' define 'text'/'question' as strings but lack minLength, maxLength, or pattern constraints.
No output schemas documented. Tools return strings (transcript, seconds, or fallback '[nothing heard]') but LLMs cannot infer structure. Rubric requires documented return types for A+ tools.
Error handling is minimal and non-actionable. 'listen' and 'ask_user' return '[nothing heard]' on failure; 'speak' returns '0.0' or '-1' without explaining why or what to do next. Rubric requires recovery guidance.
No tool annotations despite all tools being read-only. FastMCP supports readOnlyHint; adding it would clarify that these tools have no side effects and are safe to retry.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
Parameter descriptions lack constraint details. 'text' in speak should specify max length (e.g., '1-2 sentences, <200 chars') to prevent LLMs from passing oversized inputs. Rubric requires format, range, and allowed values in descriptions.