The voice cockpit for AI agents · MacDEFINE

znn-os · JARVIS is a voice cockpit for AI agents that runs on your Mac. You speak or type a request, and it hands the work to real AI agents running in real terminals, with Claude Code as the reference agent. Instead of a single chat window, you get a grid of live terminals, one per agent, and a signal that tells you which agent is waiting for your decision. It also adds local memory, a dashboard of agents, services and token use, autonomous epics, and a connection to the skills, tools and MCP servers you already have.

It is built for founders, executives, independents and developers who have more to do than days to do it, and who want to delegate to AI agents without losing sight of what those agents are doing. Developers feel at home because the terminals are real and the agent is Claude Code. Non-developers can use their voice and follow the grid and the dashboard.

Dictation apps such as Wispr Flow or Superwhisper turn your speech into text in any application, and AI launchers combine search with a chat. znn-os · JARVIS does something different: it turns speech into delegated work. The request goes to real agents in real terminals, you watch them in a grid, and you decide when one of them asks for a decision. It complements dictation tools rather than replacing them.

You need a Mac, Apple Silicon or Intel, running macOS, with Homebrew (the installer uses it to install uv if it is missing) and Claude Code installed and signed in, with the claude command available in your PATH. Chrome or Brave is recommended because they let the intro sound play automatically. The first launch installs the voice engine dependencies and downloads the XTTS-v2 voice model, about 2 GB, once. Local dictation with Whisper MLX needs Apple Silicon.

Partly, and by design. Speech synthesis runs on your Mac with XTTS-v2, and dictation runs locally with Whisper MLX on Apple Silicon. The voice engine uses about 2.4 GB of memory and stops after ten minutes without use. The agents themselves use the models you configure: with Claude Code, requests go to Anthropic exactly as they do without znn-os · JARVIS, while local models run through Ollama and stay on your machine. Remote access is off by default.

Claude Code is the reference setup. Because the terminals are real, znn-os · JARVIS can host any command-line agent or tool you can run in a shell. Local models run through Ollama, and a local router can send small jobs to local models while keeping the large ones for Claude. The MCP servers and skills in your existing configuration stay available, so you do not rebuild your stack.

The memory is SYNAPTIC-MEM, a persistent local memory for AI coding assistants with zero cloud tokens. It stores decisions, bug fixes and patterns in PostgreSQL with pgvector and finds them by meaning with semantic search. Recall is ranked with the ACT-R cognitive model: Activation = Base-level + Weight x Similarity + Spreading + Noise. Memories you use often stay active, unused ones become dormant and then forgotten, and none is ever deleted. It exposes six MCP tools and installs with one script. The repository is public: github.com/MrZzE00/MCP-Claude-mem-local.

ACT-R is a cognitive architecture developed at Carnegie Mellon University that models how human memory retrieves information. SYNAPTIC-MEM applies its activation formula to rank memories: base-level activation rises with frequent use and decays over time as a power law, similarity measures how well a memory matches your query, spreading activation lifts memories that share tags, and a little noise avoids a frozen ranking. The implementation follows ACT-R and Honda et al., HAI '25, Human-Like Remembering and Forgetting in LLM Agents.

RTK is an open-source command-line proxy that filters and summarizes command output before it reaches the model, so agents spend tokens on work instead of logs. On the author's own setup, rtk gain reports 2.4 million of 4.5 million input tokens saved, or 52.4%, over 9,927 commands. Your results depend on your commands and workflow. RTK is a third-party tool under the Apache-2.0 licence, available at rtk-ai.app and installable with brew install rtk.

skill-stack is a router that keeps a large library of skills outside the model context. Instead of loading thousands of skill descriptions up front, the agent searches a keyword index, shortlists a few candidates and loads the body of only the chosen skill at the moment it needs it. The library costs nothing until it is used, so the context stays light. Your own local skills remain available as well.

An epic is a larger piece of work split into tasks. You launch it with a single command or sentence, either in full-auto mode or with human gates. A registry tracks the run, merges stay local, and any decision that needs you waits for your answer, flagged by the amber signal in the grid. Each worker is a separate terminal you can watch.

The dashboard shows your real data: which agents are running, the state of services, and token use. It reads from your own machine and sits next to the terminal grid, so you can see the whole picture before deciding.

A mobile view is in preview. You pair a phone with a pairing code, follow one tab per agent, use a key bar with Esc, Tab, Ctrl and arrows, and revoke access with one command. Remote access is disabled by default and reaches your Mac over a private network such as Tailscale.

Speech synthesis and dictation stay on your Mac. Your agents send prompts to whichever model provider you configured, for example Anthropic for Claude Code. znn-os · JARVIS itself runs as a local server on your machine. Remote access is off by default, uses a pairing code and device cookies, and can be cut with one command.

The voice speaks French by default, and the language and the voice sample are configurable through the XTTS settings. This page and the film are in English, and the app interface is in French today.

The znn-os · JARVIS repository is private for now, and pricing is not published. The memory component, SYNAPTIC-MEM, is public on GitHub, and RTK is a third-party open-source tool. Book a demo to get access and to talk about your use case.

It installs with one idempotent command, ./install.sh, so running it again is safe, and it adapts to the configuration you already have. It is shared during a demo. The memory component installs on its own: git clone https://github.com/MrZzE00/MCP-Claude-mem-local.git, then cd MCP-Claude-mem-local and ./install.sh. The installer sets up PostgreSQL 17 with pgvector, a Python environment, the database schema, Ollama and the embedding model.

Book a demo through LinkedIn: linkedin.com/in/norbert-jeff-nadir. The author is Norbert Nadir.