Accessibility and safety
Artemis is designed around a screen reader, and around one rule: the developer decides every change.
Built for screen readers
- A voice of its own, beside Orca. Artemis speaks in an ElevenLabs voice while Orca keeps reading your screen. Without a connection or a key, the offline voice shares Orca’s speech service, so Artemis stops only its own speech, never Orca’s. One voice speaks each line, never two.
- Works from any app. Push-to-talk and spoken results work even when Artemis isn't the focused window.
- Interruptible. Say “stop” or “Hey Artemis” while it speaks, or press push-to-talk, and Artemis stops talking first, so what it was saying isn't recorded. Ctrl+Alt+M turns all listening off and on.
- Keyboard first. Approvals are Enter and Escape on a focused dialog; saying “approve” is an option only for clean changes in your project. Focus moves to the dialog when it opens and back when it closes.
- Nothing silent. Every pause, failure and completion is announced, and everything spoken is kept in a log you can review.
You approve every change
- Exact approvals. An approval covers one exact edit, or for a run, the entry file you reviewed and its arguments. Other project files are not part of a run's approval. If the files for an edit change before you decide, it is cancelled and asked again.
- A spoken "yes" never approves. With conversation mode on, only the exact word “approve” can, and only for a clean write or run inside your project, after Artemis has finished reading it to you. Anything the scan flagged, anything outside your project, and publishing stay on the keyboard. “Reject” works at any time. Conversation mode is off unless you turn it on.
- Blocked actions stay blocked by default. When a security scan finds a serious problem, focus starts on Reject and Enter alone can't approve. Overriding takes a separate shortcut on the Approve button.
- No delete tool, and every command needs your approval. The agent works only through a fixed set of tools: projects and files, proposed edits, running a Python file, security scans, page checks, opening apps, files and web pages, typing into an editor, and publishing to GitHub. On Ubuntu it can also ask to run a shell command: you approve each one on the keyboard, and it stops after 30 seconds. It has no tool for deleting files.
- Files are untrusted. Text in your project, such as a comment telling the AI to delete everything, is treated as data, not as an instruction. The agent has no delete tool, and any change it proposes still needs your approval. An approved program runs with your permissions, so read it before you approve.
Your data
- Speech stays on your computer. Recognition runs locally. Audio is held in memory only while you speak and is discarded after transcription. It is never saved or uploaded.
- The agent uses Claude. Your request, the project's file names, the files the agent reads and its tool results (such as a program's output) are sent to Anthropic's Claude API to plan changes. The API key is never put in prompts, logs or speech.
- Spoken lines use ElevenLabs. The text Artemis says aloud is sent to ElevenLabs to be voiced, with secrets removed first. Your recorded voice is not. Without an ElevenLabs key, the offline voice speaks and nothing is sent.
- Local only. The core listens on your own computer and requires a secret token for each launch, so websites can't control it.
- This site has no cookies or analytics. Its one small script only runs the visual motion and the Pause motion button, and it loads nothing from other servers.
Known limits
- Programs run with approval, a static scan, a time limit and a minimal environment. That is gated and limited, but it is not an operating-system sandbox.
- Artemis is tested on Ubuntu 24.04 with GNOME and Orca. Other screen readers and systems are not tested yet.
- Speech recognition can mishear names and symbols. Artemis tells you its plan before it changes anything, a misheard word can’t approve a change, and typing is always available.