Atomic Hermes is a macOS desktop application that packages the Hermes Agent core from Nous Research into an Electron shell with a chat window, workspace file browser, built-in terminal, and computer-use agent. It runs from /Applications, lives in the menu bar tray, autostarts on login, and ships delta auto-updates. macOS 13 or later on Apple Silicon or Intel is required.
For computer use, every screenshot is paired with native OCR (Apple Vision on macOS, Windows.Media.Ocr on Windows) that runs offline, giving the agent pixel coordinates for on-screen labels, fields, and buttons. An on-screen overlay shows when the agent is controlling the mouse and keyboard, and a session lock prevents two agents from driving the desktop at once. The same computer-use stack is published separately as an MCP server (@atomicbotai/computer-use-mcp) and a TypeScript library (@atomicbotai/computer-use).
The file browser snapshots every file before and after the agent edits it into a hidden .history directory, with no git setup. You can list past versions with timestamps, open side-by-side diffs against the current file, and restore any version; restoring takes a fresh snapshot first.
Models can be local or cloud. A one-click local mode downloads a bundled inference engine, inspects RAM and chipset, picks a model to match, and starts an OpenAI-compatible server on the machine. Cloud mode accepts keys for providers including OpenRouter, Anthropic, OpenAI, Gemini, DeepSeek, Kimi, Moonshot, MiniMax, Nous Portal, xAI, Venice, NVIDIA, Alibaba Cloud, Ollama, and Hugging Face, and you can switch between local and cloud mid-conversation.
Features
- Chat: streaming conversations with a tool-using agent, native approval modals for dangerous actions, voice-memo input, and skill slash commands
- Files: Monaco-based editing, side-by-side diffs, and automatic version history for every agent-touched file
- Terminal: a built-in PTY with background processes, streaming output, and per-command approval
- Computer use: screen control with native OCR, an activity overlay, and a session lock
- Messaging gateway: one agent reachable from Telegram, Discord, Slack, WhatsApp, Signal, iMessage, SMS, email, Matrix, Microsoft Teams, Feishu, DingTalk, BlueBubbles, Mattermost, Home Assistant, and Twilio, with shared memory across platforms
- Hermes Agent core: self-written skills after complex tasks, agent-curated cross-session memory, 40+ tools, FTS5 session search, subagent delegation, and terminal backends for local, Docker, SSH, Daytona, Singularity, and Modal
- Skills Hub: installs curated skills from agentskills.io inside the app
- MCP: the computer-use server is seeded at first launch, and other MCP servers can be added from Settings
- Cron and profiles: scheduled unattended tasks, plus isolated agent instances with separate keys, memory, skills, and gateways
- Dashboard: a local admin panel for skills, tools, memory, and messaging configuration
