Heard

A voice layer for macOS that turns terminal output from Claude Code, Codex and Cursor into smart spoken summaries

Vibe Coding speech synthesis Developer Tools 语音摘要

https://heard.app/

Visit Website

In-depth Report

Heard is a macOS speech layer designed for AI programming agent workflows. It connects Claude Code, OpenAI Codex and Cursor to convert terminal output into intelligent speech summaries. It was launched on Product Hunt on July 25, 2026 and ranked first on the day (339 votes, 91 comments). It is open source (Apache-2.0) and free for individuals. The core idea is to "broadcast with judgment" - not simply converting text to speech, but distinguishing what should be said and what should be skipped.
Kelly (@itskellysun), the founder of Heard, said this on Product Hunt: "Everyone has been trying to make it easier to talk to the agent for the past two years - voice input, better prompts, better context. But no one is doing the other half. The agent's answer is still scrolling text that you have to sit and stare at. Heard is that half." Judging from the timeline, Heard released an early version in the community as early as May 1, 2026, and then entered the field of vision of English technology media such as AI Untethered. After officially launching on Product Hunt on July 25, it received 339 votes and 91 comments, ranking first that day. At present, the product is in the early stage of rapid growth, and there is no special report by Chinese technology media (36Kr, Minority, etc.). It is worth noting that the domain name heard.app currently points to a completely different hearing aid technology product, and the actual official website entrance of Heard is currently mainly accessed through the Product Hunt page. The product uses a Python daemon with Unix socket communication, the TTS engine supports ElevenLabs (cloud) and Kokoro (local), and uses Claude Haiku 4.5 to handle personality rewriting.
The core mechanism of Heard is "judgmental reporting", which is a two-layer decision-making system. The first layer is called hard signals: permission requests, tool call failures, run completions - these always trigger voice broadcasts. The second layer is the context-aware layer: it tracks conversation history, determines what content is worth saying, and skips redundant information. The product provides three monitoring modes. Co-pilot mode gives you short prompts while you work without reading them verbatim; Companion mode provides a more complete voice briefing when you are away from the screen; Focus mode remains silent and only speaks when your intervention is required (approval, blocking, failure). Multi-agent parallel processing is Heard’s unique selling point. When multiple AI sessions (Claude Code, Codex, Cursor) are running at the same time, Heard does not have five terminals superimposed on each other, but is summarized at the project level, and each agent uses a different sound, allowing you to distinguish who is doing what by ear. The mobile phone pairing function (Heard Power) is completed through a one-time QR code. After pairing, you can listen to the broadcast in real time even when the screen is locked. It also supports pressing and holding or clicking to speak to send instructions to the smart agent. Scope pairing and automatic unpairing ensure security. The voice personality system is also very sophisticated. There are four built-in personalities: Aria (calm and direct), Friday (bright and lively), Jarvis (calm and witty, default), and Atlas (dramatic). You can also customize your personality using Markdown files. In terms of privacy, the core engine is Apache-2.0 open source and fully natively self-hosted. Session state is only stored in local memory and disk, not across the network. The mobile phone only transmits audio and does not transmit session status.
Heard has three tiers of pricing. The Free plan is permanently free and includes native Kokoro voices, built-in AI and voice keys, custom personalities and hotkeys, and is based on an open source engine. The Pro plan is $12 per month (paid annually) and offers managed voice and LLM (no configuration required yourself), all-day reporting, premium personalities (Atlas and Friday), and cross-Mac cloud sync. The Power plan costs $24 per month (paid annually) and includes all the features of Pro plus Heard mobile, code-optimized speech recognition, and voice approval and redirect agent capabilities. Hosting plans require macOS 13 or higher. This pricing strategy is more reasonable: the free version allows individual developers to try and evaluate it at low cost, the Pro version solves the needs of users who don’t want to fiddle with API configuration by themselves, and the Power version is aimed at heavy users and mobile office scenarios.
Feedback from the Product Hunt community has been overall positive. User Wes Carlson commented: "Running the agent in parallel makes the separation of hard signals and context awareness feel like a core design choice rather than just a notification feature." Dogan Akbulut said: "I always miss the moment when the agent is waiting for permission prompts, wasting 20 minutes. Love this near-silent mode." Noctis Leonard believes that the Verbosity knob and project-level summary are the right infrastructure for working in parallel. Many users also asked good questions. Gal Dayan is concerned with how Heard decides what is worth saying when several parallel agents are running different tasks. Others asked about the voice interruption mechanism, the storage location of conversation status, etc. The founder responded to these technical issues in the comment area and was quite candid.
In terms of English technology media, AI Untethered first reported on Heard on May 1, quoting the founder: "The hardest part is not TTS, but deciding what not to say." daily.dev also made an introduction after the release of PH, focusing on the technical principles of the two-layer decision-making system. Tool aggregators like AIToolly and AIDeckly provide complete features and pricing information. From an industry perspective, Heard fills an overlooked gap. Most AI programming tools focus only on “input”—how the user speaks to the agent. The question of how the agent "outputs" it to the user has never been taken seriously. It is this output problem that Heard chose to solve. In the context of the increasing popularity of AI programming agents, there is indeed a demand for this "audio-first human-computer interaction" direction. However, there are currently almost zero Chinese media reports on Heard. This may be related to the fact that the product is only spread in the English community and the official website domain name points to another product.
There are currently no significant disputes or legal risks for Heard. But there are several potential issues worth paying attention to. The first is the separation of product positioning and domain name - heard.app points to hearing aid technology rather than AI voice layer products, which may cause user confusion and also affect natural traffic acquisition. Secondly, there is competitive product pressure: Anthropic’s official Claude Code/voice mode has been launched, and open source alternatives such as VoiceMode MCP are also continuing to develop. Heard needs to establish sufficient technical barriers. The third is pricing. The $24 Power plan is not expensive in the field of AI tools, but if the main competing products are open source and free, it needs to continue to prove the unique value of paid functions.
Heard is best suited for three types of people. The first category is individual developers who use Claude Code or Codex on a daily basis, especially those who don’t want to be tied to a terminal all the time. The second category is developers who work in remote or mobile scenarios and can use their commuting and walking time to follow the progress of the agent. The third category is heavy users who run multiple agents at the same time and need project-level summary rather than switching back and forth between various terminals. People who are not suitable include: developers using Windows or Linux (not supported yet, but official cross-platform plans have been confirmed), developers who do not need voice feedback, and users who use AI programming agents less frequently. If you don’t need the native macOS experience, VoiceMode MCP or Claude Code’s built-in /voice mode can be used as an alternative.
Heard solves a real but easily overlooked problem—the output-side experience of AI-programmed agents. The product design is thought through in depth, from hard signal/context-aware two-layer decision-making to multi-agent independent timbre, and the understanding of the actual workflow can be seen in every function point. The open source core and three-tier pricing strategy also gives users flexible choices. The product release time is short (only 2 days), and subsequent developments are worthy of attention, especially cross-platform support and the promotion of the Chinese community.

User Reviews

Cynthia_RodriguezII

—

The multi-agent parallel processing feature is a lifesaver. I used to juggle four terminal windows and it drove me crazy. Now each agent speaks with a different voice, so I know who is doing what with my eyes closed.

LUphi

—

Just tried Heard's Companion mode. I left my desk to grab a coffee while listening to Codex refactoring that old project of mine in the background. When I came back the code was done and even the tests passed. What a great feeling.

TendermintTina54

—

Focus mode is my favorite. It doesn't chatter constantly — it only speaks when something errors out or needs my approval. I used to waste twenty minutes with Claude Code because I missed a permission prompt; not anymore.

Anthony.Rogers58

—

Tried the phone pairing feature — one QR scan and it was paired. You can hear announcements even with the screen locked, and even hold-to-talk to send the Agent commands. Following progress while out for a walk — technology really does change life.

Diane_Price_77

—

Open-source Apache-2.0 plus a free tier — full marks for sincerity. The local Kokoro voice isn't as natural as ElevenLabs, but being able to run it at zero cost is already a great deal.

Logan216

—

Heard turning Agent output into voice — this direction is exactly right. Everyone's busy building voice input; nobody's touching the output side. The founder is right: the Agent still answers in scrolling text, and you have to sit there staring at it.

谢秀

—

It hit #1 on PH that day with 330 votes, which shows the demand really exists. The developer tools space is brutally competitive — getting to #1 isn't easy.

梅花_15

—

Yesterday before clocking off I ran three Agents on different jobs at once — Claude Code refactoring the backend API, Codex writing front-end components, Cursor running tests. Heard announced each one's progress in three different voices: the Aria persona was especially clear on the API part, and Friday's upbeat tone handled the front-end updates — impossible to mix them up. Midway one Agent got stuck on a permission approval, and Focus mode immediately pushed a voice alert; I tapped approve and it kept going. On the way home I listened to the rest of the announcements on my phone, and when I opened my laptop at home, all three tasks were done. An experience I couldn't even have imagined before.

Nancy8552024

—

The four voice personas each have their own character. I tried Friday — the upbeat style really does suit long coding sessions. The open design of customizing personas with Markdown shows some real thought too.

DWhite_Pro509

—

The Pro plan is $12 a month, with hosted voice and LLM so you don't have to configure your own API. Pretty good value for people who don't want to fiddle, but I'll run the free tier for a while first.

goldenmouse658

—

What struck me most was the line 'the hardest part isn't TTS, it's deciding what not to say.' Subtraction is much harder than addition, and Heard really does a good job filtering out noise — it doesn't just turn all terminal output into speech, it selects with judgment.

郝玉梅

—

Read the discussion on Product Hunt — the founder's replies to technical questions were very candid. Session state lives in local memory and disk, and the phone only gets audio, never state. That privacy design is solid.

NSullivan

—

I switched over from the open-source claude_voice project. Heard's experience is far better — not even the same class of product. They've truly delivered 'announcements with judgment,' not just simple TTS reading. And Heard's architecture is more lightweight: the Python daemon plus Unix socket approach is way more elegant than running a Node service just for TTS, with a smaller footprint too.

AustinMorales168224

—

The Power plan at $24 a month is a bit pricey, but it includes mobile and voice-approval features. For someone like me who's often out and about, I genuinely need it — I'll trial it for a month and see if it's worth it. If the mobile experience is good enough, the price is actually acceptable.

Ryan_StephensJr5

—

One question: if two parallel Agents report contradictory results — one says the migration succeeded, the other says tests failed — how does Heard summarize that? Saw the founder's reply on PH: it uses a reasoning-pass override mechanism. Ship first, optimize later.

AnnMendozaII

—

Heard's Python daemon plus Unix socket architecture is pretty lightweight, with low resource usage. Much cleaner than those Electron-wrapper solutions. Using Claude Haiku 4.5 for persona rewriting is a creative touch too.

CHOKM0F

—

Heard the developers say Linux and Windows support is coming — that's great. My main machine is a Mac, but I need it to work on the occasions I'm on Windows too.

钱建晴

—

Tried it in a noisy café today and Heard still announced everything clearly. The ambient noise was a bit distracting, but the voice itself was loud enough — perfectly audible with AirPods on.

TheHelenaVergara

—

In Co-pilot mode it chimes in occasionally while you work, without breaking your train of thought. It's like a colleague sitting next to you quietly saying 'tests passed' or 'there's an error over there, take a look.' The balance is just right.

KrinHarper

—

Ran Claude Code refactoring a codebase all day with Heard — the experience exceeded expectations. Not only can I hear the progress, it also alerts me promptly when the Agent gets stuck and needs my decision. Compared with eyeballing the terminal like before, efficiency is way up. The open-source core plus three-tier pricing also gives users plenty of choice.

春雨_5

—

I've been waiting for a tool that turns Agent output into voice, and someone finally built it. My one concern is that the heard.app domain points to a hearing-assistance tech product, which is not the same thing as this Heard — that could confuse people. Hope the team sorts out the domain issue soon.

梅花40

—

If you use ElevenLabs' cloud TTS, the results are indeed very natural. But local Kokoro is good enough and saves on API costs. Trade-offs either way.

流年393

—

Self-hosting is great — everything runs locally, code and data never leave the machine. For projects with compliance requirements, this architecture matters a lot. The mobile approach of sending only audio, never state, is also reassuring.

TerryGonzalezQ8

—

There's real competitive pressure: Anthropic's official Claude Code /voice mode is already out, and VoiceMode MCP is developing too. Hope Heard keeps its technical edge and doesn't get crushed by the big players. That said, Heard does voice announcements on the output side while the official feature is voice input — different tracks, more complementary than competitive.

nty8ahuejt

—

After installing, you can switch modes right from the menu bar — no restart, no commands to type. Very Mac-like interaction design. As an Apple ecosystem user, I find it really comfortable.

8domk0e_i3j

—

One detail worth mentioning: Heard's Verbosity knob lets you fine-tune how much gets announced, from full commentary all the way down to errors only. When working in parallel, a project-level summary beats listening to five terminals talking over each other. I'd recommend every developer using Claude Code give it a try.

夏明

—

Been using Heard with Cursor for front-end work these past couple of days — feels great. I tweak styles while listening to the Agent report its status, no more switching screens back and forth. Productivity is up quite a bit, especially since I no longer have to sit and wait while tests run.

LucaKumar

—

The lack of Windows and Linux support is definitely a weak spot. Hope the cross-platform plan on the roadmap lands soon, otherwise teammates on different systems can't all standardize on it.

Heard

In-depth Report

User Reviews

Similar Products