Voice layer for Claude Code and Codex

Your coding agent has a voice now.

Heard speaks your coding agent's outputs so you can get up, walk around, and still know what's going on.
Think Jarvis for your coding agents — a voice layer for Claude Code, Codex, OpenClaw, Hermes, and any AI workflow: your agent narrates its work out loud, you talk back hands-free.

_{Pairs with — or replaces — input tools like Wispr Flow: on Power, Wispr Flow-style dictation is built in (push-to-talk + ambient voice typing, on-device speech-to-text), and Heard speaks everything your agent says back.}

_{Comparing macOS coding-agent notification tools? Heard covers the notifier basics — you hear it when Claude Code or Codex finishes, fails, or needs approval — then goes past them: text-to-speech for the work itself, narrating progress, questions, and results across parallel sessions.}

heard.dev  ·  Releases  ·  Discord  ·  Issues

See and hear it run

🔊 Turn sound on for demo

https://github.com/user-attachments/assets/d823a946-fb6f-438b-904f-aa66d4268ed1

Get the app

The app is the managed experience: download, sign in, and you're narrating - Heard runs the cloud voices and the narration brain for you, no keys to manage. Prefer to run it yourself with your own keys? See Self-host.

Have your coding tool install it (recommended)

Paste this into Claude Code, Codex, or any AI coding tool:

Install Heard so you can narrate your work to me out loud. Run: curl -L https://heard.dev/download/cc -o /tmp/heard.zip && unzip -o /tmp/heard.zip -d /Applications && xattr -dr com.apple.quarantine /Applications/Heard.app && open /Applications/Heard.app, then hand it back to me - a quick setup window opens and I'll take it from there.

Manual

Download the latest Heard.zip, drag Heard.app into /Applications, double-click to launch. Onboarding walks you through sign-in, permissions, and your voice.

Codex

Heard supports both Codex CLI and Codex App.

• Codex CLI: turn on Codex in Heard, then open Codex CLI, type /hooks, and trust the Heard hooks.
• Codex App: keep Heard running from the menu bar. Heard watches Codex Desktop's local session log and narrates new app activity automatically once Codex is enabled.

You should not need to run a development daemon. If Heard ever starts with a stale daemon socket or pid file, the app now cleans that up on launch.

Plans

	Voices	Talk back	Price
Free — self-host: build from this repo	Local only — Kokoro or your own keys, zero cloud	-	Free
Pro	All cloud voices + personas, run for you	-	$15/mo
Power	All cloud voices	Yes - hands-free voice control + built-in Wispr Flow-style dictation (push-to-talk & ambient voice typing)	$30/mo

Free is the open-source path - clone this repo and run the engine with your own ElevenLabs / Anthropic keys or fully local Kokoro; no account, nothing through our cloud. The downloadable app is the official closed build - a native successor to this engine with the cloud voices and narration brain run for you; sign in and your plan decides what's on. The app is key-free by design, so a lapsed trial just goes quiet. Power adds hands-free voice control and built-in dictation - you talk to your agent and it acts, and you can voice-type anywhere your cursor is. See pricing →

Invite a friend and you both win: they start on a 30-day trial, and you get a free week of Pro for each friend who activates.

Your Jarvis moment

Every AI workflow deserves a voice. Heard is the voice layer: a Jarvis-style narrator (that's literally our default voice's name) that turns your agent's stream of tool calls and replies into spoken updates — and on Power, listens back, so you converse with your agent like Tony Stark talks to Jarvis. First-class with Claude Code, Codex CLI, and the Codex app today — and it rides along anywhere those agents run (OpenClaw, Hermes Agent, your own stack), with heard run wrapping anything else.

Built-in dictation on Power — Wispr Flow style

Power ships with Heard Whisper, a built-in dictation and voice-typing tool in the spirit of Wispr Flow, Superwhisper, and macOS dictation — but wired straight into your coding agent:

• Push-to-talk: hold Right-⌘, speak, release — your words are transcribed and typed at your cursor, in any app.
• Ambient mode: always-on, hands-free voice input with voice-activity detection — talk naturally, interrupt Heard mid-sentence (real echo cancellation), and it types or acts as you go.
• On-device speech-to-text: transcription runs locally on your Mac — fast, accurate, private; no audio leaves your machine.
• Call-aware: joins a Zoom/Meet/Teams call and ambient input pauses itself, then resumes after.

So on Power, Heard is the full voice loop: speech-to-text in, agent does the work, text-to-speech out.

What it does

• Narrates with judgment, not just transcription. Heard decides what to say based on context - your recent activity, what tool just ran, whether something is a decision moment or routine progress. Not every tool call gets the same airtime.
• Three listening modes you switch between. Co-pilot for screen-on work - short hooks and signposts. Companion for eyes-off (driving, cooking, walking) - fuller briefings that name the choice and surface the decision. Focus for alert-only use - quiet unless something needs your attention.
• Multi-agent aware. Run 3+ agents in parallel; Heard voices the most salient one and quietly summarises the others. Each gets a distinct voice so you can tell them apart by ear.
• Talk back, hands-free (Power). Not just listening - reply out loud and your agent acts on it. Approve a step, answer a prompt, or redirect it without touching the keyboard.
• "Catch me up" (Power). Step away, come back, and ask your coding companion what you missed - Heard speaks a fresh recap of exactly the window you were gone: what landed, what's still running, what's waiting on you. "Catch me up on everything" sweeps all your active projects at once, mission by mission.
• Four personas, fork-your-own. Aria (calm, direct), Friday (bright, breezy), Jarvis (Marvel butler), Atlas (cinematic narrator).
• Works with any coding CLI. First-class adapters for Claude Code, Codex CLI, and Codex App; heard run <command> wraps anything else.

Personas

Persona	Vibe
aria	Calm, direct, never editorial. Senior pair-programmer.
friday	Bright, breezy, three steps ahead. Sprinkles "boss".
jarvis	Marvel JARVIS-coded butler. Dry wit, "Sir" only on summaries.
atlas	Cinematic narrator. Greek tragedy applied to compile cycles.

Fork your own - drop a Markdown file with frontmatter into ~/Library/Application Support/heard/personas/.

Listening modes

Switch from the menu bar → Mode.

Mode	When	What you hear
Co-pilot (default)	At the screen, coding	Short hooks and signposts. Routine tool churn gets a one-liner; decisions and finals get fuller narration. The details live in the diff you can read.
Companion	Hands-off - driving, cooking, walking	Lean but substantive briefings. State the choice, surface the decision, plain English over developer-speak, every turn ends with a hook into action.
Focus	Focused elsewhere, but reachable	Alert-only. Speaks for approvals, blockers, failures, and decisions that are waiting on you; routine progress and normal finals stay quiet.

Running multiple agents

Heard's brain handles cross-agent salience automatically - when 2+ sessions are firing, the one with the most salient signal (blocked, decision moment, failure) gets voiced; the others get summarised. Each session is given a distinct voice so you can tell them apart by ear.

Pin a specific session if you want to focus: menu bar → Active agents → click one. Click again to unpin.

Tuning

The basics - persona, voice, speed, mode, pause/resume - all live in the menu bar. Hotkeys: ⇧⌥. to pause, ⇧⌥, to resume.

Deeper knobs (verbosity profiles, per-repo overrides, narration preferences) live in Settings or .heard.yaml. Handy one: drop label: My Project in a repo's .heard.yaml and the voice announces that project by the name you chose instead of the folder name. Most users never need to touch the rest - Heard's listening modes cover the common cases on their own.

Self-host (open source)

Heard is Apache-2.0. The packaged app above is the managed experience; if you'd rather run it from source - your own keys, no account, full control - clone and configure it:

git clone https://github.com/heardlabs/heard.git
cd heard
python3 -m venv .venv && source .venv/bin/activate
pip install -e .

# bring your own keys - used directly by the daemon, nothing through our servers
heard config set elevenlabs_api_key <your-key>   # voice (skip this → local Kokoro)
heard config set speechify_api_key <your-key>    # voice, alternative (Simba 3.2)
heard config set anthropic_api_key <your-key>    # narration brain (skip → neutral templates)

# wire up your coding agent - the daemon auto-starts on the first tool call
heard install claude-code        # also: codex-cli, codex-app

That's the DIY path: you own keys, updates, and config. Everything's configurable (personas in heard/personas/*.md, verbosity in heard/profiles/*.yaml, per-repo .heard.yaml). The managed tiers are the same engine with the voices + brain run for you.

FAQ

How do I catch up on what Claude Code did while I was away?

Say "catch me up" (or "what did I miss?") and Heard speaks a fresh recap of your away window: what each agent finished, what's still running, and what needs you. "Catch me up on everything" covers all active projects at once. It re-summarizes rather than replaying old narration, so hours away come back as a few sentences - the pull half of an AI pair programmer that talks. (Power)

Does my agent's output leave my machine?

Depends on which backends you opt into.

• Voice synth. ElevenLabs and Speechify send spoken text over HTTPS. Kokoro runs fully locally - nothing leaves the machine.
• Narration. Heard sends compact event summaries (what tool ran, the agent's response text, recent context) to the Heard narration brain - a fast LLM pass that decides what to say and shapes it in your persona's voice. Either through your own Anthropic key, through Heard's managed proxy if you're signed in, or - with no key and no sign-in - falls back to neutral templates locally.

What does ElevenLabs actually cost in practice?

The free tier covers light daily use. A heavy day of pair-programming (2-3 hrs of narration) typically lands in the few-cents-to-low-dimes range on the paid Starter plan. Switch to Kokoro (free, local) for a hard ceiling.

Will narration slow down my agent?

No. Hooks fire-and-forget over a Unix socket; the daemon synthesises and plays asynchronously. Your agent never blocks on Heard.

Is this open source? How do I contribute?

Yes - Apache 2.0. The easiest places to contribute are adapters (heard/adapters/), personas (heard/personas/*.md), and verbosity profiles (heard/profiles/*.yaml).

Compatibility

macOS 14+ · Claude Code + Codex CLI/App first-class · Cursor and Aider planned · anything else via heard run.

Status

Releases on this repo are the official closed app (the download surface); this open-source engine is built from source — see Self-host. Engine status: cross-event-judgment narration via the Heard brain (each meaningful event is judged against your recent context and the active agents before anything is said). Co-pilot / Companion / Focus listening modes, multi-agent salience with a distinct per-window voice each, hands-free voice control on Power, and automatic failover across ElevenLabs / Speechify / local Kokoro. First-class Claude Code, Codex CLI, and Codex App adapters. Used daily by the author. Backward-compatible API surface; deeper knobs may move into preferences over time.

License

Apache 2.0.

Heard includes third-party speech components. Full credits and license texts are in THIRD-PARTY-NOTICES.md.