FAQ
What is Atomic Agent?
Atomic Agent is an open-source, MIT-licensed agent runtime that runs on your own machine. It drives your browser, files, shell, and git through a set of built-in tools, and the language model behind it runs locally through llama.cpp rather than a hosted API, so installing it needs no account, no API key, and no per-token billing. The runtime owns the loop: each inference returns one JSON array of tool calls constrained by a GBNF grammar, the runtime executes them, compresses the results, updates durable state, and asks the model again until the task is done. Conversations, memory, and scheduled tasks persist in SQLite files on your disk instead of growing inside the prompt, risky actions pass through an approval gate before they run, and external tools connect over the Model Context Protocol. It is a Developer Preview, currently v0.6.5, built by AtomicBot, with source at github.com/AtomicBot-ai/atomic-agent.
Is my data sent to the cloud?
Your conversations, files, and memory stay on your machine by default: the agent loop, the model, and the state directory are all local. Some things do use the network: anonymous usage analytics and crash reports (on by default, turn them off with "analytics": { "enabled": false } in config.json), the startup update check, model downloads in managed mode, and anything you explicitly set up, such as a cloud model, MCP servers, Telegram, or the browser tools.
What does it cost to run?
There’s no token bill. You bring your own llama-server (or let the CLI manage one), so the work runs where you do and the meter stops at zero, unless you choose a paid cloud provider.
What platforms are supported?
Atomic Agent is in Developer Preview (currently v0.6.5). Builds ship for:
- macOS on Apple Silicon (arm64). There is no Intel Mac build.
- Linux on x64 and arm64.
- Windows on x64 (Windows on ARM runs the x64 build under emulation).
The prebuilt binary embeds its own runtime. Running from source requires Node.js 25.7.0 or newer.
What model does it use?
Any model your local llama-server serves. The published GAIA Level 1 benchmark was run on the same model and hardware as Hermes for a fair head-to-head comparison.
How does it compare to Hermes?
On GAIA Level 1 (53 tasks, same hardware and same model), Atomic Agent scored 69.8% versus Hermes’ 58.5% — +11.3 percentage points more accurate and about 1.6× faster per task.
Is it really open source?
Yes — released under the MIT license. Open weights, open source, open traces: software you can understand all the way down. The code lives on GitHub.
Is Atomic Agent backed by NVIDIA?
Atomic Agent is built by AtomicBot, a member of NVIDIA Inception since August 2026. Inception is NVIDIA’s free program for AI startups: it takes no equity and involves no funding, so this is a technical membership, not an investment. It gives us access to NVIDIA developer resources and cloud credits.
Nothing about it changes the runtime. Atomic Agent stays MIT-licensed and local-first, and an NVIDIA GPU is not required. Managed mode uses Metal on Apple Silicon, Vulkan on Linux x64 (NVIDIA, AMD, and Intel), CUDA or Vulkan on Windows depending on your driver, and a CUDA build with CPU fallback on Linux arm64.
Can I extend it?
What hardware do I need?
The download is about 45 to 50 MB; installed, the binary is about 140 MB because it embeds its own Node runtime, plus a few small support folders next to it. The model is what needs real resources. The default qwen-3.5-4b needs a few GB of disk and enough memory to load. GPU acceleration is used automatically when available: Metal on Apple Silicon, Vulkan on Linux x64, CUDA or Vulkan on Windows, and CUDA with CPU fallback on Linux arm64. On macOS the runtime budgets 75% of unified memory; on Linux and Windows it reads your GPU’s VRAM. If a model reports as not fitting, pick a smaller one.
Which model should I use?
qwen-3.5-4b is the default and a good starting point — small, fast, and light on VRAM, which suits tool-routing and repetitive work. The catalog covers Qwen 3.5, 3.6, and 3.8, Gemma 4, Nemotron, and Muse, plus embedding models for fuzzy recall. You can also add a GGUF from any Hugging Face repo.
The catalog moves between releases, so atomic-agent models list is the source of truth for what you can actually pull. Use models pull <id> and models use <id> to switch. See Local models for the full picture.
Do I have to use a local model?
No — local is the default, but not the only option. You can point Atomic Agent at any OpenAI-compatible HTTP endpoint, or configure a cloud provider, if you’d rather. Those paths are opt-in: nothing reaches a cloud model unless you set it up. See Local models.
Where does Atomic Agent store my data?
Everything the agent persists lives under one state directory — by default ~/.atomic-agent, overridable with ATOMIC_AGENT_STATE_DIR. That folder holds your config (config.json), secrets (.env, written with mode 0600), session transcripts, and memory. Treat it like an SSH-key directory: secrets are stored in plaintext locally. See Why local-first for the full layout.
How do I update it?
Run atomic-agent update (or accept the update prompt in the TUI). atomic-agent update --check only reports whether a newer release exists, and --version <tag> installs a specific one. Re-running the install command from Installation also works. To update the model backend in managed mode, use atomic-agent models update.
How do I uninstall it?
Run atomic-agent uninstall. It lists everything it will delete (the state directory including downloaded models, the binary and its atag alias, the support folders next to it, and the PATH line the installer added) and asks you to type uninstall to confirm. Preview with --dry-run; keep your data with --keep-data. Non-interactive runs need --yes. Nothing is installed system-wide. See Uninstalling.
It’s not working — where do I start?
See the Troubleshooting page. The most common issues are the model server not being healthy (atomic-agent models status) and the wrong GPU being picked (atomic-agent models devices).