Who Atomic Agent is for#
Developers triaging GitHub issues and creating PRs
Atomic Agent clones a repository, reads an open issue, traces the cause to specific source lines, and opens a fix on a branch, pausing before it pushes. The GitHub integration uses your saved token to list issues, create pull requests, and comment, with all writes approval-gated.
Skip if:
You need the agent to push to remote repositories without an approval step, or to interact with GitHub Actions CI pipelines. Remote git sync is disabled by default and requires explicit configuration before the agent can fetch or push.
Researchers gathering and summarizing web sources
The agent drives a real browser across multiple pages and PDFs, extracts text, follows links, and writes a cited note directly into the local memory store. Notes are full-text searchable via SQLite FTS5 and can be recalled in future sessions without re-browsing the same sources.
Skip if:
Your research sources are behind authentication walls requiring human interaction or CAPTCHA-gated pages. The browser automation can handle standard logins, but adversarial anti-bot protections may interrupt the task.
Teams running AI tasks under compliance or air-gap constraints
The full agent loop, model, and data store run locally by default with nothing sent to external servers. Cloud providers are opt-in and clearly configured. This makes Atomic Agent suitable for legal, medical, or enterprise teams where data must stay on-premises.
Skip if:
Your compliance requirements extend to the model weights themselves, such as restrictions on using open-weight models released under specific terms. Verify the license of each model you choose independently.
Power users scheduling recurring monitoring tasks
Cron-style scheduled tasks let the agent run periodic jobs like market price checks, news monitoring, or score tracking, then send a Telegram notification only when something changes. The Telegram integration provides remote control from a phone, with approval buttons for risky actions.
Skip if:
You need reliable sub-minute scheduling or integration with enterprise task schedulers. The built-in scheduler suits personal or small-team workflows, not high-availability production pipelines.
The problem it solves#
Cloud-based AI agents cost money every time the model touches your data. OpenAI Operator, Manus, and similar proprietary services run on vendor infrastructure, so each task burns API credits, each session sends context to their servers, and the bill scales with usage. For compliance-sensitive workflows, sensitive documents, or air-gapped environments, a cloud-first agent simply cannot be the right tool.
Even for teams without strict privacy requirements, the economics add up. A long research task or a multi-step code triage job can consume thousands of tokens across dozens of tool calls. At commercial API rates, that is a meaningful cost per task. And when the vendor changes pricing or sunsets a feature, your workflow breaks on their timeline, not yours.
How it solves it#
GBNF grammar-constrained tool calls
Every tool call produced by the local model is constrained at generation time by a GBNF grammar, so the JSON format is structurally valid before the agent executes it. This eliminates the parse failures and format errors that break other local-model agents, and it works with any quantized open-weight model served by llama.cpp, including models too small to follow format instructions reliably on their own.
KV-cache reuse with TurboQuant llama.cpp
Atomic Agent ships its own TurboQuant fork of llama.cpp with WHT-rotated KV-cache quantization that compresses the cache up to 6.4x versus F16. Combined with a stable prompt prefix that keeps persona, tools, and instructions byte-identical across turns, cache_prompt and slot pinning mean the model re-encodes only the new context each step, keeping long multi-step tasks fast on consumer hardware.
Browser automation via Playwright
The agent drives Chrome, Edge, Brave, or Chromium through playwright-core, navigating pages, clicking, typing, and managing tabs using compact ARIA snapshots clipped to a 24k-character budget instead of expensive full-page screenshots. Risky actions pause for approval; independent read-only calls run in parallel.
MCP and Composio tool connections
External tools connect through the Model Context Protocol, so any MCP server's tools, resources, and prompts join the agent's registry. The Composio integration adds over 1,500 SaaS tools (Gmail, Slack, Notion, Linear) with OAuth handled automatically. Every write to a real account stays approval-gated.
SQLite-backed memory with hybrid recall
Memory lives outside the prompt in a local SQLite database with FTS5 full-text search and optional embedding-based recall. The agent stores profile facts, episodic notes, lessons, and procedures, then injects only compact pointers into each prompt turn. Reflection runs after turns in a separate slot so it does not block the reply.
Strengths and trade-offs#
Strengths
- MIT license with complete self-hostingThe entire agent runtime, state, and model run on your hardware under the MIT license, so you can use it commercially, fork it, and run it in an air-gapped environment with no licensing fees. Unlike OpenAI Operator (API-only, per-token billing) or Manus (cloud-dependent), your costs after setup are electricity and disk.
- 69.8% GAIA Level 1 accuracy on same model as HermesOn the GAIA validation Level 1 split (53 tasks), Atomic Agent scored 69.8% accuracy against Hermes' 58.5%, an 11.3 percentage point advantage using the same qwen-3.6-35b-a3b model and hardware. It completed tasks in roughly 217 seconds versus 351 seconds for Hermes. The benchmark is fully reproducible, with raw artifacts published in the GitHub releases.
- Fully offline and air-gapped operationThe agent loop, model inference, and all state run locally by default. No data leaves the machine unless you configure a cloud provider. This makes it suitable for compliance-sensitive workflows, offline environments, and teams who cannot send document contents to external servers.
- Import from Claude Code, Codex, Hermes, and othersOn first run, Atomic Agent offers to import skills, memory, sessions, and MCP servers from Claude Code, OpenAI Codex, Hermes, OpenClaw, Pi, and Oh-My-Pi. The import is non-destructive and does not touch the source agent's data, lowering the switching cost from existing agents.
Trade-offs
- -Developer preview: APIs and commands still changeThe README notes this is a developer preview and that APIs, commands, config, and behavior are still in flux. Teams that need a stable integration point should pin a specific release version. Changes between versions may require updating scripts or configurations that depend on the CLI interface.
- -Local model quality depends on your hardwareAccuracy improves significantly with larger models. The 35B qwen-3.6-35b-a3b reached 69.8% on GAIA L1 while the 9B qwen-3.5-9b reached 52.8%. Running the 35B model comfortably needs a GPU with enough VRAM for Q4 quantization (roughly 6.7 GB Q4 versus 24 GB for F16). Users on low-VRAM machines will need smaller models and should expect lower accuracy.
- -No macOS Intel buildsCurrent builds target macOS (Apple Silicon), Linux (x64 and arm64), and Windows (x64). Intel Macs are explicitly not yet supported; the README states this directly. Windows ARM runs the x64 build under emulation. Developers using older Intel MacBooks cannot run the CLI without building from source.
Atomic Agent vs alternatives#
Atomic Agent vs OpenAI Operator
OpenAI Operator is a cloud-based AI agent from OpenAI that completes browser tasks on managed infrastructure. Atomic Agent is an open-source, self-hosted alternative that runs the same browser automation locally on your own hardware.
| Feature | Atomic Agent | OpenAI Operator |
|---|---|---|
| License | MIT | Proprietary |
| Self-hosting | Yes | No |
| Model | Open-weight via llama.cpp | GPT-4o (managed) |
| Data handling | All local by default | Processed on OpenAI servers |
| Cost | Free (local compute only) | Subscription or API pricing |
| Offline operation | Yes | No |
Atomic Agent is the better choice when data privacy, air-gapped operation, or cost at scale are the deciding factors. Because the model and agent loop run locally, there is no per-task cost and no data leaves the machine by default. OpenAI Operator is worth considering when you need the highest-quality model performance without managing local infrastructure, or when tasks require capabilities that small open-weight models cannot match.
Atomic Agent vs Manus
Manus is a cloud-first AI agent designed to handle complex multi-step tasks on remote infrastructure. Atomic Agent covers similar territory (browser automation, file tasks, research workflows) but runs entirely on your own machine.
| Feature | Atomic Agent | Manus |
|---|---|---|
| License | MIT | Proprietary |
| Self-hosting | Yes | No |
| Infrastructure | Your hardware | Manus cloud |
| Data locality | On-device | Cloud-processed |
| Cost model | Free (local) | Subscription |
The practical difference is that Atomic Agent lets you inspect every prompt, trace every step in a local NDJSON log, and modify the runtime itself under the MIT license. Manus handles infrastructure for you, which reduces setup friction but removes the ability to audit or customize the agent's behavior. For teams who need to verify what the agent did and why, the local trace and open source codebase are a meaningful advantage over proprietary cloud platforms.
Quick start#
Install the CLI on macOS or Linux with the one-line setup command.
```bash
curl -fsSL https://atomicagent.io/install | sh
```What it's built on#
- Languages
- JavaScriptTypeScript
- Frameworks
- React
- Runtimes
- Node.js
- Tooling
- esbuildPlaywright
FAQ#
Does Atomic Agent require an OpenAI API key or cloud account?
No. Atomic Agent runs open-weight models locally via llama.cpp and requires no API key for the default setup. You bring your own model or let the CLI download one; all inference happens on your hardware. Cloud providers like OpenAI or Anthropic can be configured as optional fallbacks, but nothing is sent to external servers by default.
What operating systems and hardware does Atomic Agent support?
Atomic Agent runs on macOS (Apple Silicon), Linux (x64 and arm64), and Windows (x64). The desktop app requires macOS 14 or later on Apple Silicon, Windows x64, or Linux x64 and arm64. Intel Macs are not currently supported. The CLI requires Node.js 25.7 or later. Model quality and speed depend on available VRAM; a 35B Q4 model needs roughly 6.7 GB VRAM, while a 9B model runs on considerably less.
How accurate is Atomic Agent compared to other local AI agents?
On the GAIA Level 1 benchmark (53 real-world tasks), Atomic Agent scored 69.8% accuracy using qwen-3.6-35b-a3b via llama.cpp, compared to Hermes Agent's 58.5% on the same model and hardware. That is 11.3 percentage points better, with tasks completing in roughly 217 seconds versus 351 seconds for Hermes. Full benchmark artifacts and methodology are published in the GitHub releases under the gaia-l1-eval-2026-06-11 tag.
Can I connect Atomic Agent to external tools and APIs?
Yes. Atomic Agent connects to external tools through the Model Context Protocol (MCP) and through Composio, which adds over 1,500 SaaS integrations including Gmail, Slack, Notion, and Linear. External tool connections are configured from the Integrations tab, and all writes to real accounts remain approval-gated.
Is Atomic Agent actually open source or source-available?
Atomic Agent is released under the MIT license, confirmed in the GitHub repository. MIT is a permissive OSI-approved license; you can run it commercially, modify it, and redistribute it with no restrictions beyond attribution. The NVIDIA Inception membership noted on the website is a free startup program that involves no equity or funding and does not change the license terms.
Similar open-source tools#
rakazo
AI teammates you own: your keys, your model, your machine.
Hermes Agent
Self-hosted AI agent with persistent memory, multi-channel chat, and model choice across OpenAI, OpenRouter, and custom endpoints.
Langflow
Visual AI workflow builder that deploys as an API or MCP server.
Athena-Public
Persistent memory for your AI across IDEs
Litellm
Self-hosted AI gateway for 100+ LLMs in OpenAI format
Overmind
Turn agent production traces into fine-tuned models you own.