What is Atomic Agent?

Atomic Agent is an open-source, local-first AI agent runtime that runs entirely on your own machine. It uses llama.cpp to serve a local model, GBNF grammars to keep tool calls structurally valid, the Model Context Protocol (MCP) to connect external tools, and a SQLite-backed memory. It's built by AtomicBot and released under the MIT license on GitHub at github.com/AtomicBot-ai/atomic-agent.

Is Atomic Agent free?

Yes. Atomic Agent is free and open source under the MIT license, and there's no token bill because the model runs locally on your own hardware. You bring your own llama-server (or let the CLI manage one), so the work runs where you do and the metered cost stays at zero.

Does Atomic Agent run offline, and is my data private?

Yes — by default the agent loop, the model, and your data all run locally on your own machine, so it works without sending your data to the cloud. Cloud providers are used only if you deliberately configure them. This makes Atomic Agent a strong fit for private, air-gapped, or compliance-sensitive workflows.

How is Atomic Agent different from Hermes?

On the GAIA Level 1 benchmark (53 tasks, run on the same model and the same hardware for a fair head-to-head), Atomic Agent scored 69.8% versus Hermes' 58.5% — that's +11.3 percentage points more accurate, solving 37 of 53 tasks. It was also about 1.6x faster per task (roughly 217 seconds versus 351 seconds on average). Unlike cloud-first agents, Atomic Agent is local-first and open source, so you keep both the accuracy and full control of your data.

What models does Atomic Agent support?

Atomic Agent works with any model your local llama-server can serve via llama.cpp, so you're free to choose the open-weight model that fits your hardware and task. GBNF grammars constrain the model's output so tool calls stay valid regardless of which model you pick. The published GAIA Level 1 benchmark was run on the same model and hardware as Hermes to keep the comparison fair.

What operating systems does Atomic Agent support?

Atomic Agent is currently in Developer Preview (v0.1.65) and runs on macOS, Linux, and Windows today. You can follow progress and see current requirements on GitHub at github.com/AtomicBot-ai/atomic-agent.

Is Atomic Agent really open source?

Yes. Atomic Agent is released under the permissive MIT license, with open weights, open source, and open traces — software you can read and audit all the way down. The full code is on GitHub at github.com/AtomicBot-ai/atomic-agent.

Can I extend Atomic Agent with my own tools?

Yes. You can connect external tools through the Model Context Protocol (MCP), teach it repeatable workflows with skills, and even drive it remotely from Telegram. Because it's MIT-licensed and open source, you can also modify the runtime itself.

Local First AI Agent

It runs on your machine, drives your real tools, and keeps every session on your own disk.

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

📁 sosidudku — atomic-agent — 143×45

            -:::::::--
            -::::::::-
           -:::::::::-
          -::::::::::-
         -:::::::::::-
       -:::::::::::::-
    -::::::::::::::::-
-::::::::::::::::::::::::::::::::-
::::::::::::::::::::::::::::::::::
::::::::::::::::::::::::::::::::::
-:::::::::::::::::::::::::::::::::
=------------:::::::::::::::::---=
 @@@@@@@@@@@*-::::::::::::-=+#%%@
            -:::::::::::-+#@
            -::::::::::=#@
            -:::::::::=#
            -::::::::-*
            -::::::::=
            +--------*
              %%%%%%

▄▀█ ▀█▀ █▀█ █▀▄▀█ █ █▀▀   ▄▀█ █▀▀ █▀▀ █▄░█ ▀█▀
█▀█ ░█░ █▄█ █░▀░█ █ █▄▄   █▀█ █▄█ ██▄ █░▀█ ░█░

local operator agent — Ink + llama.cpp

• Entersubmit message to the agent
• /helplist all slash commands
• /sessionsswitch to a previous thread
• /newstart a fresh session
• /observe · /manage · /runoperator dashboard (Tab cycles panels in debug)
• Ctrl+C ×2quit (once aborts a running turn)

❯ █

● healthy · 2 ms · qwen-3.5-9b · llama.cpp

[enter] send [alt+enter] newline [tab] sidebar [ctrl+b] open Observe [/] commands [ctrl+c] quit ·[shift] select

Join Community Documentation

Atomic Agent Hermes Agent

GAIA Level 1 – 53 real-world tasks on the same model and same machine

Accuracy · GAIA Level 1

Atomic Agent

★ Winner 69.8%

Hermes Agent

58.5%

+11.3pp More accurate

37/53 solved. Grammar-constrained decoding (GBNF) forces structurally-valid tool calls; bounded tails keep reasoning on-budget.

1.6× Faster per task

217s vs 351s. cache_prompt + slot_id reuse across steps, plus parallel read-only batches that overlap I/O after a single inference.

15 vs 9 Often right

Grammar-locked actions and step-by-step checks mean Atomic Agent rarely walks into an unrecoverable wrong answer.

how it works

↻ the loop repeats – KV-cache keeps each step cheap, nothing re-sent to a server

/01

You ask

A task, in plain words — from the terminal or Telegram.

/02

Local model decides

One pass on your own GPU returns a list of next actions — grammar-locked, always valid.

/03

It runs them

Browser, files and shell — in parallel. Pauses for approval on risky steps.

/04

Repeat → done

Loops with cached context, so each step stays fast — until it replies or you stop it.

features

the engineering that keeps it fast, private, and running on your own machine

KV-CACHE

A byte-stable prompt prefix means every turn reuses the same cached compute. Nothing is re-encoded, and nothing leaves your machine.

prefix  ▓▓▓▓▓▓▓▓▓▓  cachedtail    ░░░  re-computed────────────────────step N  KV-cache HIT  ✓

GBNF

Grammar-constrained decoding forces structurally valid actions, even from tiny local models.

root ::= "[" call     ("," call){0,15} "]"call ::= { "tool":…, "args":{…} }

TURBOQUANT

Curated Q4 GGUF quants fit a long window on an everyday GPU.

fp16  ████████  ~24 GBq4    ██        ~6.7 GB

SQLITE MEMORY

Recall injected in the tail, a few hundred tokens, cache intact.

SELECT content  FROM memories_fts  ORDER BY rank→ ~320 tokens

LLAMA.CPP

It pulls the backend and model for you. No accounts, no bills.

$ atomic-agent models startchat: started, healthy● on port 8080

Connected
to 0+ apps

Bring your daily tools into one assistant

Use cases

files · sort

INV

IMG

ZIP

Scanning ~/Downloads · 318 files
Reading docs for real titles
Grouping → /Invoices /Shots
Renamed 212 · deduped 7

files, folders, done Sorts, renames, groups, and cleans your files — reading each one to understand it first.

cron · watch

BTC/USD64,210 ▲

NEWS llama.cpp v0.5 ships

MATCHARS 1–1 CHE

→ @you

Cron · every 30m
BTC $63,420 (+2.1%)
Arsenal 2–1 · 78’
Pinged Telegram ✓

cron in, Telegram out Schedule a recurring check — markets, news, match scores — and get a Telegram ping only when something changes.

git · triage

acme/api main

cloning…

#42 NullRef on init

PR #128

Cloning acme/api
Reading issue #214
Found in loader.go:88
PR fix-214 ready ✓

issue in, PR out Clones a repo, reads an issue, finds the cause in the code, and opens a fix on a branch — pausing before it pushes.

browse · note

wikiarxiv.org

PDF

note.md source: arxiv ✓ saved to memory

Browsing 6 sources
Parsing 2 PDFs
Extracting · 12 cites
Note saved · searchable ✓

browse, read, save Drives a real browser across sources, reads PDFs and pages, and writes a cited note into your local memory.

Run Agent

any model, local or cloud

Run local agent
in one click

Local-first agent in one app. Same engine, zero setup. Join the waitlist to get in first!

EMAIL ADDRESS

macOS • Linux • Windows

FAQ

Everything you need to know about running a local-first agent.

What is Atomic Agent?

Atomic Agent is an open-source, local-first AI agent runtime that runs entirely on your own machine. It uses llama.cpp to serve a local model, GBNF grammars to keep tool calls structurally valid, the Model Context Protocol (MCP) to connect external tools, and a SQLite-backed memory. It's built by AtomicBot and released under the MIT license on GitHub at github.com/AtomicBot-ai/atomic-agent.
Is Atomic Agent free?

Yes. Atomic Agent is free and open source under the MIT license, and there's no token bill because the model runs locally on your own hardware. You bring your own llama-server (or let the CLI manage one), so the work runs where you do and the metered cost stays at zero.
Does Atomic Agent run offline, and is my data private?

Yes — by default the agent loop, the model, and your data all run locally on your own machine, so it works without sending your data to the cloud. Cloud providers are used only if you deliberately configure them. This makes Atomic Agent a strong fit for private, air-gapped, or compliance-sensitive workflows.
How is Atomic Agent different from Hermes?

On the GAIA Level 1 benchmark (53 tasks, run on the same model and the same hardware for a fair head-to-head), Atomic Agent scored 69.8% versus Hermes' 58.5% — that's +11.3 percentage points more accurate, solving 37 of 53 tasks. It was also about 1.6x faster per task (roughly 217 seconds versus 351 seconds on average). Unlike cloud-first agents, Atomic Agent is local-first and open source, so you keep both the accuracy and full control of your data.
What models does Atomic Agent support?

Atomic Agent works with any model your local llama-server can serve via llama.cpp, so you're free to choose the open-weight model that fits your hardware and task. GBNF grammars constrain the model's output so tool calls stay valid regardless of which model you pick. The published GAIA Level 1 benchmark was run on the same model and hardware as Hermes to keep the comparison fair.
What operating systems does Atomic Agent support?

Atomic Agent is currently in Developer Preview (v0.1.65) and runs on macOS, Linux, and Windows today. You can follow progress and see current requirements on GitHub at github.com/AtomicBot-ai/atomic-agent.
Is Atomic Agent really open source?

Yes. Atomic Agent is released under the permissive MIT license, with open weights, open source, and open traces — software you can read and audit all the way down. The full code is on GitHub at github.com/AtomicBot-ai/atomic-agent.
Can I extend Atomic Agent with my own tools?

Yes. You can connect external tools through the Model Context Protocol (MCP), teach it repeatable workflows with skills, and even drive it remotely from Telegram. Because it's MIT-licensed and open source, you can also modify the runtime itself.