Not a chatbot — a real AI assistant that lives on your desktop. Chat, terminal, file browser with time-travel snapshots, a built-in computer-use agent, free local models, and a mission control for 16+ messengers — all in one beautiful, self-contained app.
Atomic Hermes is a native macOS AI assistant — not a browser tab, not a CLI wrapper, not "ChatGPT with buttons". It's an autonomous agent with hands, eyes, memory, and a real workspace. Launch it from /Applications, live in the tray, autostart on login, get delta auto-updates. Everything below is one app, one install, one window.
- Chat — streaming conversations with a tool-using AI agent, native approval modals for dangerous actions, voice-memo input, skill slash-commands.
- Files — a real workspace browser with Monaco-powered editing, side-by-side diffs, and automatic version history for every file the AI agent touches. Manage your skils and agent memory
- Terminal — built-in PTY where the AI agent can run commands for you: background processes, streaming output, per-command approval.
- Computer Use — the AI agent literally sees and operates your Mac, with native OCR so it clicks the right pixel the first time.
- Dashboard — a local admin panel for configuring the agent's skills, tools, memory, messaging gateway and etc.
- Settings — one-click switch between cloud AI provides and a fully local model engine we bundle for you.
Most "AI computer use" tools downscale your 2560×1600 screen to something the model can swallow, then make the agent guess coordinates. Small text turns to mush. Buttons get missed. Retries burn tokens.
Atomic Hermes does it properly. Every screenshot the AI agent takes is paired with native OCR — Apple Vision on macOS, Windows.Media.Ocr on Windows — fully offline, zero API keys. The agent gets back pixel-accurate coordinates for every label, field, and button on screen, so it clicks the right thing the first time — fewer retries, fewer wrong clicks, lower bills.
A visual "AI agent active" overlay keeps you aware whenever the assistant is driving your mouse and keyboard, and a session lock stops two agents from fighting over the desktop.
Pre-wired into the Atomic Hermes AI agent out of the box. Also open source for anyone to use in their own agent:
@atomicbotai/computer-use-mcp— MCP server, drop it into Claude Desktop, Cursor, Windsurf, or any MCP client@atomicbotai/computer-use— the core TypeScript library: OCR, actions, overlay, session lock
This is the feature that makes trusting an AI agent with your real project actually feasible.
Atomic Hermes silently saves a snapshot of every file before and after the AI agent edits it — into a hidden .history directory next to your workspace. No git required, no setup, no manual anything.
- Watch the AI agent evolve your code. Open any file in the Files tab and the right-hand panel lists every past version with relative timestamps — "just now", "12m ago", "3d ago" — plus file size.
- Diff anything against anything. Click a snapshot to open a side-by-side diff against the current file. See exactly what the agent changed, line by line.
- One-click restore. Any past version, any time. Restoring itself takes a fresh snapshot first, so there's never a dead end.
If the AI agent wrecks a file, it's one click away from being fine again. That's the deal.
Want an AI agent that lives entirely on your machine, with no keys, no API bills, and no data ever leaving the Mac? One click in the app and Atomic Hermes does the whole thing for you: downloads a bundled inference engine, scans your hardware (RAM, chipset), picks a model that will actually run well, and starts an OpenAI-compatible server locally. No terminal. No config files. No "figure out which quant fits your GPU". Every feature — chat, computer use, tools, memory — works identically to cloud mode.
Or bring keys for 20+ cloud AI providers: OpenRouter, Anthropic, OpenAI, Gemini, DeepSeek, Kimi, Moonshot, MiniMax, Nous Portal, xAI, z.ai, Venice, NVIDIA, Alibaba Cloud, Xiaomi MiMo, Ollama, HuggingFace, and more. Switch the model powering your AI agent between local and cloud mid-conversation — no restart, no config reload.
Turn on the messaging gateway from the dashboard and the same AI assistant picks up conversations from everywhere you already chat — one agent, one process, one shared memory, all of these platforms:
Voice memo transcription, cross-platform conversation continuity, shared slash commands. Ask the AI agent something from your phone on the train and find the answer — and the side-effects — waiting on your Mac when you get home.
Under the desktop shell beats the full Hermes Agent core by Nous Research — an open-source, self-improving AI agent with one of the most serious tool-calling loops around. It's what turns Atomic Hermes from a chat window into an AI assistant that actually gets smarter the more you use it:
- Self-improving skills — the AI agent writes its own procedures after finishing complex tasks; next time you ask, it already knows how.
- Agent-curated memory — the assistant decides what's worth remembering across sessions, with periodic nudges to persist knowledge. Dialectic user modeling via Honcho.
- 40+ tools — the AI agent can use all of them: file ops, web search/extract, code execution, MCP client, subagent delegation, cron scheduling, browser automation, and more.
- Full-text session search — FTS5 + LLM-summarized recall across every conversation you've ever had with the agent.
- Subagents & parallelism — your AI assistant can spawn isolated sub-agents that tackle workstreams in parallel; multi-step pipelines collapse into zero-context-cost turns.
- Six terminal backends — local, Docker, SSH, Daytona, Singularity, Modal. Run the same AI agent on a $5 VPS or a GPU cluster.
- Prompt caching done right — the agent never breaks its own cache mid-conversation, so long sessions stay cheap.
We track upstream closely and contribute back. If you want the headless CLI agent, Linux / WSL / Termux installs, RL training environments, or the raw Python agent core, the upstream repo is where to go.
- Skills Hub — browse and install curated skills from agentskills.io inside the app and the AI agent gains new abilities on the spot (GitHub, Figma, Notion, Obsidian, Trello, Google, Web Search, media, and more).
- MCP-native — the
computer-useMCP server is seeded for you at first launch; plug in any other MCP server from Settings and the agent auto-discovers its tools. - Cron — tell the AI assistant "every Monday at 9am, summarize last week's PRs to Slack" and it runs unattended.
- Profiles — run multiple fully isolated AI agent instances (personal, work, research), each with its own keys, memory, skills, and gateway.
- Approval & guardrails — whenever the agent is about to run a dangerous shell command or write to a sensitive file, a native modal asks you first. You stay in control.
- Download the latest release for macOS and drag into
/Applications. - Launch Atomic Hermes, pick an AI model for the agent to use (cloud key or free local), start chatting.
- Hit the Computer button when you want the AI agent to operate your Mac. Open Files to watch the assistant's edits accumulate in real time as snapshots.
- Optional: flip on the Messaging Gateway from Settings → Messaging and talk to the same AI agent from Telegram / Discord / Slack / anything.
macOS 13+ on Apple Silicon or Intel. Auto-updates ship deltas so subsequent versions are tiny.
Prerequisites:
- macOS 13+ (Apple Silicon or Intel)
- Node.js ≥ 18
- Python 3.11 (installed automatically via
uvon first build) - Xcode Command Line Tools (
xcode-select --install) - uv, rg (ripgrep), optionally ffmpeg
git clone https://github.com/AtomicBot-ai/atomic-hermes.git
cd atomic-hermes
cd desktop
npm install
npm run devnpm run dev will:
- Rebuild native modules against your Electron version (
node-pty). - Run
scripts/bundle-dev.sh— symlinks the Python AI-agent sources intodesktop/build/, creates auv-managed venv, and links system binaries so edits to the agent core are picked up instantly. - Build the admin dashboard (
../web) and the renderer (Vite). - Compile TypeScript in watch mode and launch Electron.
The desktop app ships with a self-contained Python 3.11, a uv-managed venv with all agent deps, the Hermes agent source, ripgrep / Node / ffmpeg, and the built-in skills — everything lives under desktop/build/. Two scripts manage that tree:
npm run bundle:dev # lightweight dev bundle — symlinks repo sources into
# desktop/build/, reuses your existing venv. Called
# automatically by `npm run dev`; safe to re-run (~2s).
npm run bundle # full production bundle — wipes desktop/build/,
# downloads relocatable Python 3.11 via uv, runs
# `uv sync --all-extras`, copies binaries, installs
# browser node_modules. Required before `npm run dist`.npm run build:all # full TS + dashboard + renderer build (no bundle)
npm run dist # full bundle + signed & notarized .zip in desktop/release/
npm run dist:local # same, but without signing (for local smoke tests)
npm run release:patch # tag, push, trigger GitHub release workflowFull release flow: npm run bundle && npm run dist. Signing & notarization happen via the electron-builder hooks in desktop/scripts/ and expect Apple Developer credentials in the environment.
atomic-hermes/
├── desktop/ # Electron AI agent desktop app (this is what ships)
│ ├── src/main/ # Electron main process (IPC, windows, updaters)
│ ├── src/python-server/ # FastAPI bridge between Electron and the AI agent
│ ├── renderer/src/ # React UI (chat, files, terminal, settings, setup)
│ ├── scripts/ # dev bundler, release, afterSign hooks
│ └── package.json
├── agent/ tools/ gateway/ # Hermes AI agent Python core (imported by desktop)
├── hermes_cli/ # Headless CLI agent entrypoints & config wizard
├── cron/ # Scheduler for unattended AI agent jobs
├── skills/ optional-skills/ # Built-in skill library the agent can load
├── web/ # Admin dashboard (served inside the desktop app)
└── website/ # Docs site (Next.js)
Issues and PRs welcome. For desktop-specific work, edit files under desktop/. For AI agent core work (adding tools, changing the reasoning loop, editing the tool registry, the skin engine, or profile-safe code patterns), follow AGENTS.md at the repo root — it covers the full Python agent architecture.
Please run the Python test suite before opening a PR:
source venv/bin/activate
python -m pytest tests/ -q- Website — atomicbot.ai/hermes
- Releases — github.com/AtomicBot-ai/atomic-hermes/releases
- Issues — github.com/AtomicBot-ai/atomic-hermes/issues
- Computer Use MCP — @atomicbotai/computer-use-mcp
- Computer Use Library — @atomicbotai/computer-use
- Upstream — NousResearch/hermes-agent
- Privacy — atomicbot.ai/privacy-policy
- Support — support@atomicbot.ai
The Atomic Hermes AI agent desktop distribution is licensed under PolyForm Noncommercial 1.0.0. The underlying Hermes Agent core is MIT-licensed by Nous Research.