Dodaios (formerly Personal Jarvis) is an open-source, local-first desktop AI assistant, officially described as an "AI assistant for the agent era" and a "desktop agent operating system." It sits between the user and AI agents, responsible for running agents, maintaining memory for agents, connecting tools, and keeping the user in control. The project is licensed under Apache 2.0 (versions 1.6.0 and earlier retain the MIT license), supports Windows, macOS, and Linux, and can also run as a screenless server.
In terms of use, Dodaios aims to cover all kinds of things AI can do on a computer with a single desktop app: after the user states a need, it can answer directly, perform tasks on the computer, or hand the task to the appropriate agent, and let the user view every step. It does not require an account, and the official cloud does not intervene between the user and the model.
Main capabilities include:
- Agent IDE: Splits the screen into multiple panes, each running a coding agent, and assigns call signs such as T1, T2, etc., which the user can direct by voice, for example "have T1 write tests," "what is T2 doing," "send the fix plan to all panes." It can discover installed coding tools and launch them in a pane with one click, and supports sending the same task to multiple panes at once and comparing the results. Each pane is a real terminal, and the user can manually type at any time.
- Agent teams: Expert agents such as researchers, release managers, and marketing writers can be created with one click. Each agent has a never-ending chat that can continue across days; agents keep private notes and learn from their own work; routine tasks can be set to run automatically on a schedule; produced files, reports, and plans go into Artifacts and show the tools used. Longer coding tasks run in a git worktree copy of the repository, can be split into parallel work units, and go through review before the user views the results.
- Voice interaction: Users can customize the wake word (such as "Hey Nova" or "Computer"), engage in natural turn-taking conversation, and mid-way ask to open a website, check the calendar, or start an agent. Dictation can input speech into any application and learns the user's vocabulary through a personal dictionary. Phone calls are made through the user's own number, and calls can be answered or, after approval, made outbound. Speech recognition and speech can run fully offline, with hosted speech as an optional feature.
- Tool connections: Plugins can be installed with one click from the plugin marketplace, or users can add MCP servers themselves; it can discover GitHub, Vercel, Supabase, Stripe, Docker, and cloud CLIs, etc.command-line tool, and log in within the app; Skills are reusable instructions that are automatically enabled when a request matches; accessible via Telegram and Discord; the jarvis command and local API give user scripts and other agents the same level of control.
- Security and control: every operation goes through a four-level risk check (safe, monitor, ask, block), operations that may change the system require user approval, and all runs are logged. Keys are stored in the operating system's credential store, and data does not leave the local machine unless a connection requires an external service.
How to use: can be installed via a one-line install command (with corresponding commands for Windows PowerShell, macOS, and Linux); the installer checks for Python 3.11+ and Git, prompts installation if missing, adds a desktop launcher, and opens the app; running it again updates the existing installation. You can also download the desktop installer (macOS Apple Silicon, Windows installer), or use pip install personal-jarvis and jarvis serve to run on a server, then open in the browser http://localhost:47821。从另一台电脑使用麦克风需要 HTTPS. After installation, complete a brief setup within the app: choose a language, a wake word (or keyboard shortcut), and a model.
Use cases: need a local, open-source, controllable desktop AI assistant; want to run multiple coding agents simultaneously and command them in a unified way; want to control the computer, browser, and apps by voice; want the agent to have long-term memory and routine tasks; want to connect existing CLI tools and MCP services; want to use an existing Claude/ChatGPT subscription or local models without paying extra.
Limitations and requirements: requires one of an API key, a subscription, or a local model; voice requires a microphone; no GPU is required. The macOS disk image has not yet been notarized by Apple; on macOS 15, the first open requires clicking "Open Anyway" in System Settings, and on macOS 14 and below, you need to right-click to open; the one-line installer does not have this step. When manually installing with pip or pipx on an Intel Mac or Windows ARM64, a native package index is also required. Supported models include OpenAI, Anthropic Claude, Google Gemini and Vertex AI, OpenRouter, NVIDIA, Ollama, and any OpenAI-compatible local server, as well as Claude Code and Codex subscriptions; coding agents include Claude Code, Codex, OpenCode, Kimi Code, GLM, Grok Build, Antigravity, Cursor CLI, and DeepSeek Harness; speech recognition supports local Whisper, OpenAI, Gemini, Groq, Deepgram, and OpenRouter, and speech synthesis supports Piper (local), ElevenLabs, Cartesia, Inworld, Gemini, and OpenRouter.
