Jarvis Desktop: a voice assistant powered by your AI subscriptions for Mac, Windows and Linux

⚡ The gist in 5 seconds - The gist: the desktop assistant Jarvis is activated by the voice command "Jarvis" and performs tasks through AI subscriptions the user is already paying for — ChatGPT Plus/Pro, Claude Code or Codex. - Where it's available: an open-source project on GitHub (lopadova/jarvis-desktop), version v0.1.4 in pre-release status; installers for macOS, Windows and Linux are not yet signed. - Limitation: the release is early, so a one-time "open anyway" step is required; the code is open under the MIT license, and the authors explicitly invite auditing of the security rules. ### 🔍 What was found The lopadova/jarvis-desktop project is a cross-platform app built on Tauri 2: installers are only 10–15 MB, with target idle memory consumption under 80 MB. The assistant lives "next to the clock": it answers quick questions out loud, sets timers and reminders (working offline, without AI), translates and summarizes copied text, supports dictation, screen content explanation (opt-in), file search, background web research, shopping lists and editable memory. Wake word and speech recognition (Whisper) run locally; audio only goes to the cloud if a cloud voice provider is selected, and a visible indicator shows when the microphone is active. There is a fully local mode: Ollama + local Whisper + Piper — nothing leaves the computer. Sign-in is via a ChatGPT Plus/Pro subscription without an API key, or via Claude Code or Codex login. The user sets a weekly limit in their ChatGPT account: when it's reached, Jarvis reports it and never silently switches to pay-per-use API. The assistant connects as an extension to Claude Desktop / Cowork, via a one-click free relay to ChatGPT, and as an MCP server to Claude Code and Codex. Voices: ElevenLabs (including expressive tags like [happy] or [sigh]), Fish Audio, OpenAI, local Kokoro/Piper and the system voice. Commands in Italian are supported: "Jarvis, ricordami tra 20 minuti di chiamare Marco". ### 💡 Why it matters The practical benefit is a single voice control point for the desktop for those already paying for multiple neural networks, with no unexpected bills. Security is built on a risk model: background agents ask Jarvis before running commands, deleting files or sending data; low-risk requests are approved by voice ("yes"), high-risk ones require a click, and "always allow" is never offered for them. Links and web pages cannot initiate actions. ### 🧩 Context The repository is distributed under the MIT license and covered by 600+ automated tests, including tests for every security rule. The current release v0.1.4 is a pre-release, installers are unsigned, so on macOS/Windows a one-time system permission will be required. The authors ask users to open issues when problems arise: early feedback shapes the next release. According to the comparison table in the repository, Jarvis fills the niche where built-in OS assistants and chat apps can't work with an already-paid subscription via voice wake word, while CLI-only agents

🔗 Read on github.com