Heard

Heard is a macOS voice companion for AI coding agents, built by Heard Labs and released under the Apache 2.0 license. It listens to agent events from tools like Claude Code and Codex, decides what information is worth surfacing, and converts it into natural spoken narration — so developers can leave their desk without falling behind on agent progress.

What It Is

Heard sits in the macOS menu bar and wires into coding agent hooks (Claude Code, Codex CLI, Codex App). Rather than transcribing every tool call, it applies contextual judgment — using a lightweight LLM call (Claude Haiku) — to decide what to say and how to say it. The result is concise, persona-driven narration rather than a raw log readout. The core engine is open source (Apache 2.0) and can be self-hosted with your own API keys; a managed cloud tier handles voices and the narration brain without any setup.

Three Listening Modes

Heard ships with three switchable modes accessible from the menu bar:

• Co-pilot — short hooks and signposts for screen-on work; routine tool churn gets a one-liner, decisions get fuller narration.
• Companion — fuller briefings designed for eyes-off situations like walking, driving, or cooking; surfaces decisions in plain English.
• Focus — alert-only; stays silent unless an approval, blocker, failure, or decision is waiting.

No prompt tuning or restart is required to switch between modes.

Multi-Agent and Mobile Workflow

Heard handles parallel agent sessions by routing the most salient signal to voice and rolling the rest into compact background summaries. Each agent session is assigned a distinct voice so users can tell them apart by ear. The Power tier adds Heard Mobile: pair a phone once via a one-time QR code, and live narration follows to the phone with the screen locked. Users can hold or tap to talk back — instructions are transcribed on-device using NVIDIA's Parakeet TDT 0.6B v3 model and submitted to the agent running on the Mac.

Open-Source Deployment Model

The README describes Heard as Apache 2.0 open source. Self-hosting requires cloning the repo, installing Python dependencies, and supplying your own ElevenLabs (or local Kokoro TTS) and Anthropic keys. With no keys at all, Heard falls back to neutral local templates. The managed app tiers run the same engine with cloud voices and the narration brain handled by Heard's infrastructure. Custom personas can be added by dropping a Markdown file into ~/Library/Application Support/heard/personas/.

Update: v1.1.21

The GitHub repository shows the latest release as v1.1.21, published July 12, 2026. The README describes the current status as v1.0.x-era cross-event-judgment narration, Co-pilot/Companion/Focus listening modes, multi-agent salience routing, and automatic ElevenLabs ↔ Kokoro failover. The project was created in April 2026 and has seen active development, with the repository last updated in late July 2026. Cursor and Aider adapters are listed as planned additions beyond the current first-class Claude Code and Codex support.