Heard vs. Crier: two ways to hear Claude Code and Codex progress aloud.

(1)Crier

What Crier does well.

Crier’s README lists four moments when it speaks: the turn finished, the agent needs permission, something broke, or it has been waiting. By default it also says a short line when the agent automatically compacts its context. It asks the agent to end each reply with a [say] line, so the model that did the work writes the sentence, with no second API call. Permission requests use the plain description the agent already wrote next to the command, such as deleting the build folder, rather than the raw command. It never reads the full answer, and it stops mid-word when you start typing.

A bell rings before each sentence. When several projects talk at once, one speaks at a time. Permission requests and errors skip the cooldown and wait their turn rather than being dropped, while any other sentence that waits in the queue longer than 45 seconds becomes a chime. You can give each project its own voice and have it say its name when the speaker changes; both are off by default. Speech comes from a local Supertonic server on 127.0.0.1, with no network calls at runtime and no telemetry. crier doctor checks each link from config to audio player and tells you what to fix.

• Free and open source (MIT).
• Local speech, no API key, no account.
• Korean, English and Japanese.
• Claude Code, Codex CLI, Gemini CLI and Cursor.

curl -fsSL https://raw.githubusercontent.com/pg-Parunson/crier/main/bootstrap.sh | sh
# not on Claude Code? add -s -- codex, -s -- gemini or -s -- cursor to the end
# restart your agent session, then:
crier demo

(2)Heard

What Heard does differently.

Heard installs hooks for Claude Code and Codex and decides what deserves speech. Questions, approvals and failures come through as a sentence or two in every listening mode. Results do too in Co-pilot and Companion, and routine tool calls get a one-liner or stay quiet; Focus keeps routine progress and results quiet. In the app, with several sessions running, routine updates are batched into short summaries that start with the project name.

The app adds managed voices and spoken recaps on Pro. On Power you hold right ⌘ to dictate a reply into the right session, ask Parrot questions about a repo, and follow along on the iPhone companion. The open-source engine, Apache 2.0, is a separate build you run yourself with local Kokoro voices.

• Listening modes: Co-pilot, Companion and Focus.
• Summaries across sessions, project name first.
• Codex CLI and the Codex app.
• Voice replies and iPhone on Power.

(3)Where Crier wins

Pick Crier for a free, local sentence at the moments that matter.

If you want one short line when an agent is done or stuck, spoken on your Mac with no account and no key, Crier does exactly that and costs nothing. It speaks Korean and Japanese as well as English, covers Gemini CLI and Cursor turn ends and permission requests alongside Claude Code and Codex, and installs with one curl command.

(4)Where Heard wins

Pick Heard to choose how much you hear, and to answer back.

If you want progress as well as alerts, Heard’s listening modes let you hear more while you pair with an agent and go quiet when you need to focus. It covers the Codex desktop app, batches several sessions into short summaries in the app, and on Power you answer an agent’s question by voice without going back to its window.

(5)Both

Running them together.

Both add their own hooks next to existing ones, so if you set up the same agent in both you would hear two voices about the same moment. Pick one tool per agent. Crier’s installer wires one agent at a time, which makes splitting them easy: for example, Heard for Claude Code and Codex, and Crier for Gemini CLI.