walkai-talkai

walkai-talkai is a macOS application that turns any handheld two-way radio into a push-to-talk microphone and speaker for an AI coding agent. You press the PTT key on a walkie-talkie, speak your request out loud, and the app transcribes your voice on the Mac, types the command into the agent you already use, and broadcasts the agent's answer back through the radio. It is built for developers who want to talk to Cursor, Claude Code, Antigravity or a terminal-based agent without sitting at the keyboard — the radio becomes the interface and the Mac does the rest. The product describes itself as a push-to-talk voice bridge between any two-way radio and AI coding agents, running on Apple Silicon Macs with macOS 13 or later, with low latency. The product addresses a simple gap: an AI coding agent runs on your Mac, but you normally have to be at the Mac to talk to it. walkai-talkai separates the two. The radio's audio-out feeds the Mac through a simple cable, and the agent's reply is synthesized and played into the radio's VOX line, so the developer can step away from the keyboard and still issue commands and hear results. The website describes the recommended setup as leaving one radio at the Mac and taking the other with you, and its Accessibility automation is noted as typing into the exact window you chose even while you are away from the keyboard. Everything in the loop is organized around that separation — the radio is the microphone and the speaker, while the Mac handles transcription, dispatch and speech synthesis. walkai-talkai offers two operating modes so you can either drive your IDE or skip it entirely. In Target Window mode you pick a target window once, and every transmission is typed straight into that window — Cursor, Antigravity, Terminal or any app — via macOS Accessibility and then submitted with Enter, with the agent replying over HTTP. In API Mode the app talks to models directly instead of typing into an app, covering OpenAI, Anthropic, Gemini, DeepSeek and Ollama. Both modes share the same radio loop, so moving between them is a mode switch in the app rather than a different hardware setup. That means a developer can point the radio at whichever agent or model is already part of their workflow, from a local Ollama model to a hosted provider, without changing anything about the radio. Speech recognition is handled by OpenAI Whisper running locally on Metal, in Turkish and English, with live waveform monitoring in the app. Because transcription happens on the Mac rather than in the cloud, your voice is converted to text locally, and the product explicitly states that your radio audio never leaves your Mac. On the return path, reply speech is synthesized with Kokoro and EMA neural voice models running on CoreML and the Apple Neural Engine, which the site describes as fast, high-quality radio announcements. Neural voice is available in both English and Turkish. The combination means a full round trip — your spoken command in, the agent's answer out — can be completed without sending audio to an external service. Radio signalling is what makes that loop work over a real handheld. A rising 880 → 1100 → 1400 Hz preamble opens your radio's VOX before the first syllable, so the opening of the reply is not clipped, and a roger beep marks the end of the transmission. Any agent that can run curl can talk back through a one-line agent API: a POST to 127.0.0.1:1988/api/reply carrying a short text payload. That endpoint is loopback-only, so nothing is exposed to your network. Accessibility automation types into the exact window you chose, even while you are away from the keyboard. Inside the app, the station UI is organized into Station, Dashboard, Audio I/O, Automation, API Settings, Command Logs and Diagnostics, with a live spectrum that shows incoming audio in green and your reply in orange, and each message displayed underneath for a few seconds. The full loop runs in five documented stages. First, key up: press PTT and speak, and the radio's audio-out feeds the Mac through a simple cable. Second, transcribe: Whisper converts your voice to text on the Mac, in Turkish or English. Third, dispatch: the command is typed into Cursor, Antigravity or Terminal and sent, tagged as a radio command. Fourth, reply: your agent POSTs a short summary to 127.0.0.1:1988/api/reply. Fifth, broadcast: CoreML neural TTS speaks that summary into your radio's VOX line with wake tones and a roger beep. The site summarizes the architecture as Whisper, CoreML, ANE and 127.0.0.1:1988 — the radio is the microphone and the speaker, and the Mac handles everything in between. The outcomes the product describes follow directly from that design. Commands are spoken and answers are heard, which is what the site calls hands-free coding: no keyboard in the middle of the loop, and no need to return to the desk to see a result. Because recognition and synthesis both run locally, radio audio stays on the machine, and the app keeps working offline for up to 14 days between license checks. The loop is deliberately small — the site frames it as everything the loop needs and nothing it doesn't — so there is no extra cloud service sitting between you and your agent. A one-line curl request is the whole integration surface on the agent side, which means agents that already exist can talk back without modification. The concrete scenarios on the site show how this plays out. In a Target Window setup, you key up and say the buy button doesn't do anything when clicked, the phrase is typed into the agent as a radio command, and the agent answers over the radio that it checked every page, that the fix is saved and ready to ship when you say so. With two radios — one at the Mac, one carried with you — you can stay in another room and keep the loop going. In API Mode you skip the IDE entirely and call OpenAI, Anthropic, Gemini, DeepSeek or Ollama directly, receiving the model's answer as a radio broadcast. Turkish and English are both supported for recognition and neural voice, so the same station works for either language. The hardware list is short and specific: any handheld two-way radio with VOX and an external audio port — Baofeng, Quansheng, Motorola, Kenwood and similar — a standard 2-pin K-plug or 3.5 mm audio cable into the Mac or a USB audio adapter, a second radio to carry with you, and an Apple Silicon Mac running macOS 13 or later. Pricing is a single lifetime license, listed at $19.99 as an early bird price against a $29.99 regular price, one-time and excluding tax, with sales tax or VAT added at checkout and Stripe acting as Merchant of Record. One license activates one Mac and can be moved to a new Mac anytime, and free updates are included. In short, walkai-talkai turns a handheld radio into a push-to-talk remote control for an AI coding agent: local Whisper transcription in, neural voice with VOX tones and a roger beep out, with a loopback HTTP endpoint connecting the two. Its value proposition is the combination of hands-free voice control, local audio processing and a one-time license on a Mac you already own.