Feature Request: Native Voice Interaction (STT + TTS) in Claude Code

Feature Request: Native Voice Interaction (STT + TTS) in Claude Code

Summary

Add native speech-to-text (input) and text-to-speech (output) capabilities to Claude Code, enabling users to have real-time voice conversations with Claude while retaining full access to the local filesystem, terminal, MCP servers, and all existing Claude Code tooling.

The Problem

Claude Code is the most capable AI engineering assistant available today — but it's locked behind a text-only interface. This creates a fundamental interaction bottleneck: users who want to collaborate with Claude (brainstorm architecture, walk through code, discuss design decisions) are forced to type everything, even when the interaction is conversational in nature.

Meanwhile, the user's hands are often occupied — they're sketching on a whiteboard, reading documentation on another screen, working in a physical lab, or simply thinking out loud. The keyboard becomes a barrier to the natural flow of human-AI collaboration.

The gap no current tool fills

Users today face an impossible tradeoff:

| Tool | Voice? | Local File Access? | Full Tool Use? |
|------|--------|--------------------|----------------|
| Claude Code | No | Yes | Yes |
| Claude Desktop | No (yet) | Via MCP | Limited |
| Pipecat + Claude API | Yes | No | No |
| VoiceMode MCP | Partial (WSL2 only) | Yes | Yes |
| OpenClaw | Yes | Yes | Yes |

There is no first-party way to talk to Claude while it also has access to your files, terminal, and development environment. Every workaround either sacrifices file access (Pipecat creates a separate API session), requires fragile platform-specific hacks (VoiceMode MCP needs WSL2 on Windows with known audio issues), or pushes users toward third-party tools.

Why This Matters for Adoption

OpenClaw is exploding in popularity — and voice is a core reason

OpenClaw has seen massive adoption growth, and a major driver is its seamless voice interaction. Users report that being able to talk to their AI coding assistant fundamentally changes the relationship from "tool I query" to "collaborator I work with." The viral demos that drive OpenClaw adoption almost always feature voice interaction — it's visceral, it's impressive, and it makes AI-assisted development feel like the future.

**Claude Code is technically superior in capability — but OpenClaw feels more advanced because you can talk to it.** This is a perception gap that directly impacts adoption. Users choose the tool that feels like collaboration, not the one with the better benchmark scores.

Voice unlocks entirely new use cases

Today, Claude Code is primarily a coding tool. Voice interaction transforms it into a general-purpose engineering collaboration partner:

1. Model-Based Systems Engineering (MBSE): Engineers discuss system architecture verbally while Claude modifies SysML models, generates documentation, and updates design files — all in real time.

1. Research & Knowledge Work: Researchers talk through hypotheses while Claude searches codebases, updates lab notebooks, and runs analyses. The conversational flow is essential — typing fragments the thinking process.

1. Pair Programming (actual pairing): Today's "pair programming with AI" is really "solo programming with an AI chat window." Voice makes it actual pairing — you talk through the problem together while both looking at code.

1. Accessibility: Voice interaction is not a luxury for users with RSI, motor impairments, or other conditions that make sustained typing painful. This is an accessibility gap.

1. Education & Mentoring: Students and junior engineers learning a codebase can ask questions conversationally while Claude walks them through code, edits files to demonstrate concepts, and responds naturally.

The numbers case

• Users who adopt voice-first AI workflows report 2-3x faster iteration on design/architecture tasks (where most time is spent thinking, not typing)
• Voice reduces the friction to start an interaction — lowering the threshold from "let me type out a well-formed prompt" to "hey, what if we..."
• This increases session frequency, session length, and user retention — the metrics that matter for subscription products

Proposed Implementation

Claude Code's architecture is well-suited for this. The core loop is already: user input → Claude processes → tool calls + text output → user sees result. Voice just changes the I/O layer:

[Microphone] → STT engine → existing Claude Code input pipeline
existing Claude Code output → TTS engine → [Speaker]

Suggested approach

1. Modular STT/TTS providers via the existing MCP-like plugin architecture:

• STT options: Deepgram (fast, accurate, free tier), Whisper (local/private), Google STT
• TTS options: ElevenLabs (natural), Edge TTS (free), Kokoro (local)
• Let users choose based on their latency/privacy/cost preferences

1. Push-to-talk and continuous modes: Some users want walkie-talkie style, others want always-listening. Support both.

1. Voice activity detection (VAD): Use Silero VAD or similar to handle turn-taking naturally without manual push-to-talk.

1. Graceful degradation: Voice is an optional layer. Text input always works. If STT/TTS fails, fall back to text seamlessly.

1. Platform support: This should work on macOS, Linux, AND Windows natively — not behind WSL2 or other platform-specific hacks. Windows is a massive portion of the developer market.

The North Star: Full-Duplex Voice + Tool Use

NVIDIA's recently open-sourced PersonaPlex (January 2026) demonstrates where voice AI interaction is heading. PersonaPlex is a 7B full-duplex speech-to-speech model that:

• Listens and speaks simultaneously — no turn-taking delays, natural interruptions and backchanneling
• Switches speakers in 0.07 seconds — 18x faster than Gemini Live on latency benchmarks
• Scores 3.90 on Dialog Naturalness MOS (vs. Gemini Live at 3.72)
• Supports persona control via text prompts + voice conditioning

PersonaPlex proves that natural, real-time voice conversation with AI is a solved problem at the model level. But PersonaPlex is a standalone LLM — it can't edit your files, run your tests, or manage your git workflow.

The massive opportunity for Claude Code is combining PersonaPlex-level conversational naturalness with Claude's unmatched tool use and reasoning capabilities. No one has done this yet. The first AI coding assistant that lets you have a natural, full-duplex voice conversation while it simultaneously edits code, runs commands, and manages your development environment will define the next generation of developer tools.

Anthropic is uniquely positioned here: Claude's reasoning and tool use are best-in-class, and the modular architecture of Claude Code (MCP servers, extensible I/O) is already designed for this kind of integration. Whether that means integrating a model like PersonaPlex as the voice layer, building a native solution, or exposing hooks for third-party STT/TTS — the end result is what matters: talk to Claude, and Claude talks back, while never losing access to the full power of Claude Code.

What Success Looks Like

A user opens Claude Code, says "Hey Claude, let's look at the authentication module — I think we need to refactor how tokens are refreshed," and Claude responds verbally while simultaneously opening the relevant files, suggesting changes, and asking clarifying questions. The user can approve edits by voice ("yeah, that looks good, go ahead") or switch to keyboard for precise input when needed. The conversation feels natural — Claude can be interrupted mid-sentence, can backchannel ("mm-hmm, I see the issue"), and transitions seamlessly between voice and text when precision matters.

This is not science fiction — the components exist today. PersonaPlex proves full-duplex voice is solved. Claude Code proves AI tool use is solved. The integration is what's missing, and Claude Code's modular architecture is already designed for exactly this kind of extensibility.

Environment

• OS: Windows 11 (but this should be cross-platform)
• Claude Code version: Latest (via VSCode extension)
• Use case: Engineering collaboration (MBSE, knowledge engineering, research)