Best AI Agent Skills for Voice and Speech (2026)

Claude Code has had voice built in since March 2026. Type /voice, hold space, talk. Most people searching for a voice skill do not need one.

The useful question is narrower than "what is the best voice skill." It is: which of the three specific holes in the built-in feature are you trying to fill? There are exactly three, they are documented, and each one points at a different tool.

Start with what ships

/voice enables dictation in the Claude Code CLI. Your speech is transcribed live into the prompt input, so you can mix typing and talking in the same message.

Two modes. Hold mode is push-to-talk: hold Space while speaking, release to finalize. Tap mode toggles: tap once to start, tap again to stop and send. Set it per session with /voice hold or /voice tap, or make it permanent in your user settings with a voice object containing enabled and mode.

Some details that change how it feels in practice:

• Transcription is tuned for coding vocabulary. Terms like regex, OAuth, JSON, and localhost come through correctly, and your project name and current git branch are added as recognition hints automatically.
• Recording stops on its own after 15 seconds of silence or two minutes total.
• In tap mode, transcripts shorter than three words are inserted but not submitted, so a stray tap does not fire off a message.
• Transcription does not consume Claude messages or tokens and does not count toward the limits in /usage.
• Twenty dictation languages are supported, including Indonesian, Japanese, Hindi, and Turkish. It follows the same language setting that controls Claude's response language, and falls back to English with a warning if yours is not on the list.
• The push-to-talk key is rebindable. It is the voice:pushToTalk action in the Chat context, editable in ~/.claude/keybindings.json, and a modifier combination like meta+k starts recording on the first keypress with no warmup.

That is a complete dictation feature. Now the holes.

Hole one: it needs a Claude.ai account

The speech-to-text service is available only when you authenticate with a Claude.ai account. It is not available when Claude Code is configured to use an Anthropic API key directly, Amazon Bedrock, Google Cloud's Agent Platform, or Microsoft Foundry. Organizations with HIPAA compliance enabled see Voice mode is disabled by your organization's policy.

If you are in one of those situations, no setting will turn it on. You need a separate voice layer.

Hole two: audio leaves your machine

The documentation is unambiguous: voice dictation streams your recorded audio to Anthropic's servers for transcription, and audio is not processed locally. For most people that is a non-issue. For regulated work, air-gapped environments, or anyone whose threat model includes "do not send audio anywhere," it is disqualifying.

Hole three: it only listens

Dictation is one-way. The agent does not speak back. If you want an actual conversation, or you want your agent to read results aloud while you look at something else, that is a different feature entirely.

The tools that fill them

VoiceMode, the direct answer

If you want two-way voice conversation with Claude Code, this is the project built for it.

Install: claude plugin install voicemode@voicemode

Maintained by mbailey, MIT licensed, around 1.3k stars. It handles both halves: speech to text and text to speech, with cloud and local options on each side. Cloud path uses OpenAI's Whisper API and TTS. Local path uses Whisper.cpp for transcription and Kokoro for synthesis, and the installer sets these up optionally. The project describes switching between them seamlessly, and an OpenAI API key is a fallback rather than a requirement.

That combination means it addresses all three holes at once. It does not care how your Claude Code is authenticated, it can run entirely on your machine, and it talks back. If you only try one thing from this page, try this.

Set expectations on the local path. Running Whisper and a TTS model locally means real CPU or GPU load and noticeably more latency than a cloud API. The project's own framing is low latency with smart silence detection, which is achievable, but "local, fast, and high quality" is a pick-two situation on modest hardware.

ElevenLabs and MiniMax, for output quality

ElevenLabs MCP is the official server from ElevenLabs. It connects an agent to their text-to-speech and audio processing APIs, needs an ElevenLabs API key, and the free tier gives 10k credits per month. Configuration is a standard MCP server entry running uvx elevenlabs-mcp with the key supplied as an environment variable. MIT licensed.

MiniMax MCP is the equivalent from MiniMax and covers a broader media surface than voice alone.

Both are the right choice when the output is a deliverable rather than a convenience. Generating narration for a video, producing an audio version of an article, building something a listener will hear. They are the wrong choice for "read my terminal output to me," where VoiceMode's local Kokoro path is cheaper and does not bill per character.

Two honest notes. These are MCP servers, not skills, so they install through MCP configuration rather than a plugin command. And credit-based pricing for generated audio moves; check the current page rather than trusting a number in any article, including this one.

Replicate, for reaching a specific model

Replicate MCP lets an agent run models hosted on Replicate, Whisper among them. It is the general-purpose escape hatch: when you need one specific transcription or audio model rather than a whole voice stack, this reaches it without you standing anything up.

LocalAI, if nothing may leave the building

LocalAI from mudler is an open-source inference engine that runs models on your own hardware, and it runs on CPU without a GPU. Around 48k stars, MIT.

It is not a voice skill and it will not install into Claude Code as one. It is the substrate you put underneath when the requirement is that no audio, and no text derived from audio, touches a third-party server. Pair it with VoiceMode's local services and you have an offline voice loop. This overlaps with our guide to local and self-hosted AI, which covers the full picture.

Voicebox and Vexa, adjacent but useful

Voicebox is a desktop voice synthesis studio rather than an agent integration. It shows up in searches for voice skills and is worth knowing about if your actual job is producing voice audio, but it is not something your agent calls.

Vexa sends a bot into meetings and produces real-time transcripts through an API. Different problem entirely from dictation, and the right shape if what you want is "everything said in this call, available to an agent afterwards." Apache 2.0, around 2.7k stars.

Where the built-in does not work

Worth listing separately, because these catch people out and no skill on this page changes some of them.

Voice dictation needs a local microphone. It does not work in remote environments, which includes Claude Code on the web and SSH sessions. The VS Code extension supports it with the same Claude.ai account requirement, but not in VS Code Remote, Dev Containers, or Codespaces, because the microphone is on your machine and the extension runs on the host.

On WSL you need WSLg, which comes with WSL2 installed from the Microsoft Store. On WSL1 it will not work and the guidance is to run Claude Code in native Windows instead.

On Linux, if the native audio module cannot load, Claude Code falls back to arecord from ALSA utils or rec from SoX. On WSL specifically, SoX needs its PulseAudio backend installed explicitly, because WSLg routes audio through PulseAudio and the ALSA backend has no /dev/snd device to record from.

The SSH and remote limitation is a hardware fact, not a software one. Audio has to be captured somewhere. If you develop over SSH, run the voice layer on your local machine and drive the remote session from there.

Which one

• You want to dictate prompts, you are signed in with a Claude.ai account, you are on a local machine. Use /voice. Install nothing.
• You are on an API key, Bedrock, Vertex, or Foundry. VoiceMode.
• Audio may not leave your machine. VoiceMode with local Whisper.cpp and Kokoro, on LocalAI if you want the models local too.
• You want the agent to speak back to you. VoiceMode for utility, ElevenLabs MCP for anything an audience will hear.
• You are building a voice product rather than improving your own workflow. Pipecat.
• You want meeting transcripts an agent can use. Vexa.

A note on what these repositories are

A recurring problem across every agent skills directory, ours included, is that "skill" gets applied to things that are not skills. On this page: VoiceMode is a plugin, Pipecat is a Python framework, three entries are MCP servers, LocalAI is an inference engine, and Voicebox is a desktop application. Only some of them install with a plugin command.

Check the repository for a SKILL.md, or for an explicit install line in the README, before you plan around how something installs. It takes thirty seconds and it is the difference between a working setup and an afternoon of confusion.

If you are looking for voice products rather than agent plumbing, we track dictation and speech tools separately, and there are roundups for podcast hosts and content creators.