What is Atomic Agent?

A local-first, MIT-licensed AI agent that runs its control loop and all state (config, memory, sessions, traces, models) on your own machine via llama.cpp. It drives the browser, reads/edits files, runs approved shell commands, remembers context across sessions, schedules follow-ups, and calls external tools over MCP. It runs on macOS, Linux, and Windows.

How did Atomic Agent do on the GAIA benchmark?

On the public GAIA validation Level 1 split (53 tasks), Atomic Agent scored 69.8% (37/53) versus Hermes' 58.5% (31/53) — a +11.3 point gap. Both agents drove the same local Qwen3.6-35B-A3B (4-bit UD-Q4_K_XL via llama.cpp) on the same Apple M4 Max, with the same step budget and timeout. The only variable was the agent loop. Atomic also averaged ~217s/task vs Hermes' ~351s, about 1.6× faster.

Can Atomic Agent import my existing OpenClaw setup?

Yes. The first run offers to bring over your data from OpenClaw, Hermes, Claude Code, or Codex — skills, memory, MCP servers, sessions, cron jobs, and (opt-in) provider keys — with a dry-run preview before anything is written. You can also run /import in the TUI or 'atomic-agent import openclaw' later.

Does Atomic Agent only run local models?

No. It's llama.cpp-first for local models (Qwen, Gemma, Llama, DeepSeek), but also connects to cloud models — Claude, GPT, Gemini, and OpenRouter. You choose per task.

Can I control Atomic Agent from my phone?

Yes. You can message the agent over Telegram or Discord and approve its actions right in the chat, similar to OpenClaw's remote-control model.

Atomic Agent: A Local-First AI Agent That Beat Hermes on GAIA — and Imports Your OpenClaw Setup

Q: How do I install Atomic Agent?

One line. macOS/Linux: curl -fsSL https://atomicagent.io/install | sh — Windows (PowerShell): irm https://atomicagent.io/install.ps1 | iex. The installer verifies checksums, installs the CLI (aliased as 'atag'), and the tool self-updates in place via 'atomic-agent update'.

By Prahlad Menon Published 2026-09-10 4 min read

Most “local AI agent” projects are really cloud agents with a local skin — the model or the orchestration still phones home. Atomic Agent takes the harder path: the control loop and all state — config, memory, sessions, tasks, traces, downloaded models — live on your machine, and it’s llama.cpp-first so small quantized models stay useful for long, multi-step work on consumer hardware.

It’s MIT-licensed, runs on macOS, Linux, and Windows, and has picked up 2,500+ stars and nine releases in about three weeks. It’s also a member of the NVIDIA Inception program. But the reason it’s worth a look is the benchmark.

The headline: it beat Hermes on GAIA with the same model

The interesting claim here isn’t “our model is better” — both agents ran the same model. On the public GAIA Level 1 validation split (53 real-world tasks), Atomic Agent and Hermes each drove a local Qwen3.6-35B-A3B in 4-bit (UD-Q4_K_XL via llama.cpp) on the same Apple M4 Max (40-core GPU), with the same step budget and timeout. The only variable was the agent loop itself.

Metric	Atomic Agent	Hermes
Accuracy	37/53 = 69.8%	31/53 = 58.5%
Avg wall time / task	~217 s	~351 s
Head-to-head wins	+15 atomic-only	+9 Hermes-only

That’s +11.3 percentage points of accuracy and roughly 1.6× faster, from loop design rather than model size. It’s being described as the first open-source local agent to surpass Hermes on GAIA. The full methodology and per-task breakdown are published, which is the right way to make a claim like this.

Two engineering details do a lot of the work: stable-prefix caching keeps long sessions cheap, and TurboQuant shrinks the KV-cache ~6.4× so a 35B model can sustain long multi-step runs on a laptop.

Install is one line

macOS / Linux:

curl -fsSL https://atomicagent.io/install | sh

Windows (PowerShell):

irm https://atomicagent.io/install.ps1 | iex

The installer verifies the checksum, installs the CLI plus support assets (grammars, native prebuilds, bundled ripgrep), and drops a short atag alias next to the binary. It self-updates in place (atomic-agent update), and there’s a clean atomic-agent uninstall that shows exactly what it will delete before you type the word to confirm — nothing uploaded, nothing kept.

This is a developer preview, so pin a release if you need a stable integration point.

Why this one matters if you already run OpenClaw

The part that stood out to me: the first run offers to import your existing agent setup — from OpenClaw, Hermes, Claude Code, or Codex. It scans your machine, shows what it can bring over — skills, memory, MCP servers, sessions, cron jobs, and (opt-in) provider keys — and lets you confirm with a dry-run preview before anything is written. Later you can run /import in the TUI or atomic-agent import openclaw from the shell.

That’s a low-friction way to A/B a different agent loop against your current one without rebuilding your whole tool and memory stack.

Other pieces that round it out:

Phone control — message the agent over Telegram or Discord and approve its actions right in the chat (a model OpenClaw users will recognize).
Local and cloud models — Qwen, Gemma, Llama, DeepSeek locally; Claude, GPT, Gemini, and OpenRouter in the cloud, chosen per task.
1,000+ skills and MCP servers — mail, calendar, files, and browser covered out of the box, plus your own.
~1,500 apps via Composio — one API key in the Integrations hub connects them.
Coding modes — plan, auto, or ask-first, so you set how much freedom the agent gets per task.
Tunable context window — pick how many past tasks the agent carries into each request, with the cost shown live as you adjust.
Redesigned TUI — scroll, click, select, copy all work with the mouse; links open in your browser; the sidebar folds away.
Embeddable — drive it in your own apps over HTTP or a Tauri sidecar.

A desktop app is coming, with an early-access waitlist, but the terminal app is ready today.

The takeaway

Atomic Agent is a useful data point in an argument the local-agent space keeps having: how much of an agent’s quality comes from the model versus the harness around it. Holding the model, hardware, and budgets fixed and still moving GAIA Level 1 accuracy from 58.5% to 69.8% is a clean demonstration that the loop is where a lot of the value lives — the same lesson OpenClaw’s own harness work keeps surfacing.

If you run local models and already have an OpenClaw or Claude Code setup, the import flow makes it nearly free to try. Worst case, you uninstall and lose nothing.

Links: