Self-host Understudy

Understudy is an open-source, local-first AI agent you install on your own computer that carries out tasks across the whole desktop: it sees the screen and drives native apps, automates browsers, runs shell commands, searches the web, and replies over eight built-in messaging channels such as Telegram, Slack, Discord and Signal. Beyond following one-off instructions, you can teach it a workflow by demonstrating it once — it extracts intent rather than coordinates and publishes a reusable skill that survives UI redesigns — and repeated daily use can crystallize successful paths into skills automatically. Larger automations compose as workspace artifacts: playbooks sequence deterministic workers and agentic skills, for example a six-stage pipeline that browses the App Store, installs an app on a real iPhone and publishes a narrated review video. It is bring-your-own-model, with Anthropic, OpenAI, Google/Gemini and MiniMax among the supported providers, configured with API keys or OAuth. Self-hosting is through npm (GitHub Packages also works): install the package, run the setup wizard, and start the gateway daemon that terminal, dashboard, webchat and messaging surfaces connect to. Full native GUI automation requires macOS with Accessibility and Screen Recording permissions granted to your terminal app; Linux and Windows backends are planned. Layers 1–2 of its five-layer design are implemented, layers 3–4 partially, and proactive autonomy remains the long-term goal.