Atomic Agent: Pick it and plan

This page helps you decide whether Atomic Agent fits you, and which model and budget to plan for, before you install anything. It is the first of eight Atomic Agent pages, and the next one walks you through the install.

This guide is for you if you chose Atomic Agent instead of Agent Zero. Atomic Agent is a free agent made by a company called AtomicBot, and its website is atomicagent.io.

Atomic Agent is a normal app on your real computer. Agent Zero runs inside a container, but Atomic Agent does not, so when you say yes to Atomic Agent, the action happens on your real files, as you.

Atomic Agent came second of the six agents this guide compared, with 68.5 points out of 100, while Agent Zero got 70.0. Atomic Agent scored better on install (15 out of 25) but lost points on safety (12 out of 20), because a yes is not fenced in. The full scoring is in the landscape chapter.

Pick this if

• You want an agent that asks before risky actions from the start. Out of the box it runs at approval level 1, the strictest of five levels, so writing a file, running a command, and moving a file to the Trash all stop and wait for your yes.
• You understand that a yes still runs as you. The project's own security page says the approval prompt "is not a sandbox," and a command you approve runs with your own user rights.
• You are fine with a developer preview. The current version is v0.7.0, from 9 November 2026, and the docs say commands, settings, and behavior still change between releases, so screens may not match this guide.
• You have a Windows 10 or 11 PC, or an Apple Silicon Mac (M1 or later) on macOS 14 Sonoma or newer. Windows PCs with an ARM chip have their own desktop installer.
• You do not have an Intel Mac. The README says Intel Macs are not supported yet, and there is no Intel Mac download.

Before you start

The Atomic Agent program is free, but the model is separate. Read Pick your model and budget if you have not already.

By default, Atomic Agent runs a local model through a program called llama.cpp, which needs no key and no account. The first start can use a tiny starter model (about 1.3 GB), and then the app offers to download a bigger one that fits your computer.

That sounds great, but be careful. On an 8 GB or 16 GB laptop, the models that fit are weak at jobs with many steps, where each step depends on the last one. The reasons and numbers are in the local models chapter. Atomic Agent's own examples pick Qwen 3.5 4B for an 8 GB Mac and Qwen 3.5 9B for a 16 GB Mac, and the main guide rates models that size 1 or 2 out of 5 for this kind of work.

The stronger start is a free Google Gemini key, which Atomic Agent lists as a built-in model choice. You do not need a card for a free Gemini key, but Google may use free-tier prompts to improve its products.

Use your everyday account on the computer, not an administrator account.

Models it supports and what they cost

Atomic Agent can use almost any model, and nothing leaves your computer for a cloud model until you add one yourself. Prices change often, so this page names the choices and leaves the numbers to the full price table in Pick your model and budget.

• A local model, the default. It runs through llama.cpp on your own computer, so it costs nothing per use, but it needs RAM and is weak at long jobs on a small laptop.
• Google Gemini. A free key works with no card on file, which is why this guide starts there.
• OpenRouter. One key reaches many models from many companies, and you pay for what you use, so give that key a credit limit.
• Ready-made presets. The model docs list presets for Anthropic, Groq, DeepSeek, Mistral, xAI, Together, and others, plus Ollama and LM Studio on your own computer. Any company that offers an OpenAI-compatible API can be added with its web address.
• A Claude or ChatGPT plan you already pay for. Atomic Agent can run on that plan through the company's own command line tool, Claude Code or Codex, after you sign in to it. Those requests count against your plan's limits, and the shared page on subscription sign-ins explains the risks.

Once you add a cloud provider, every prompt sent to it, including any memory and file contents the agent adds, goes to that company under its own rules. The model docs say to keep sensitive work on a local model.

The catches

• A yes runs as you. The approval prompt is not a sandbox, so an approved command can reach any file your account can reach. Add-ons and skills can still read the agent's secrets file. Since v0.6.11, the commands it runs no longer get your keys, although the security page has not caught up.
• Web reading does not ask. Search and page reading run without a prompt, and the tools page says ordinary browsing does too. The docs disagree on whether raw web requests ask.
• Scheduled tasks can run with no approval. Read the warning on the phone and schedules page before you schedule anything.
• It is a developer preview. The terminal version is v0.7.0 and the desktop app is 0.0.8, and screens, names, and settings change between releases.
• No Intel Macs. The README says Intel Macs are not supported yet. The install page mentions a build-it-yourself path that needs a very new Node.js, which is not a beginner job.
• Small local models are weak. The default local path is free, but on an 8 GB or 16 GB laptop it will struggle with multi-step jobs.
• Keys are saved as plain text. They live in the .atomic-agent folder in your home folder.
• Usage reports are on by default. You can switch them off in Settings.
• Its own test result is its own. The README reports a score on a public test where Atomic Agent beat Hermes Agent on the same local model, but nobody else re-ran that test for this guide.
• Some good news. When the agent deletes a file with its own tool, the file goes to the Trash or Recycle Bin, and that asks first. Before it overwrites a file it did not make, it keeps a copy of the old one, although that does not cover a shell command you approve.
• Uninstall is total. The atomic-agent uninstall command deletes your settings, chats, memory, and downloaded models, and there is no backup.
• Some things this guide could not confirm: a desktop screen for the approval level, a desktop button for http.approvalMode, a Telegram screen in the desktop app, which Gemini models the free tier allows in Atomic Agent's list, whether Edge is picked automatically when Chrome is missing, and whether schedules run while the computer sleeps.

Sources

Checked 14 November 2026.

• Atomic Agent research notes.
• Landscape and scores.
• Local models by hardware.
• Atomic Agent GitHub repository.
• Atomic Agent README (Intel Mac note, test result, provider list).
• Latest terminal release, v0.7.0.
• Download page (Windows 10 or 11 and macOS 14 requirements, Intel Mac note).
• Docs home.
• Install and update (plain-text keys, usage reports, uninstall, build from source).
• Approval gates (the five levels).
• Tools (what asks and what does not).
• Models and cloud providers (starter model, quality picks, provider kinds and presets, prompts leave your computer).
• Connect a subscription (Claude Code and Codex plans, plan limits).
• Security page (not a sandbox, Trash, usage reports; still says skills and commands see the secrets file).
• Release v0.6.11 (commands the agent runs no longer get your keys).

Next

Next: Install and first job.