Atomic Agent is an open-source, local-first AI agent runtime built by AtomicBot, released under the MIT License on GitHub (github.com/AtomicBot-ai/atomic-agent). It runs entirely on the user's own machine and by default does not send data to the cloud, making it suitable for privacy-sensitive, air-gapped, or highly compliance-constrained workflows.
Purpose: As a local AI operator, it drives the user's real tools—browser, files, shell, and git—to execute natural language tasks issued from the terminal or Telegram. It emphasizes local operation, no keys, and no token bills.
Main capabilities:
- Local model driven: Uses llama.cpp to serve local models, supporting any open-weight model that a local llama-server can serve. The CLI can automatically pull the backend and models, with no account or billing required.
- Grammar-constrained decoding: Uses GBNF grammar to enforce valid tool-call structure, so that even small local models can output legal actions.
- Acceleration and caching: Uses cache_prompt and slot_id for reuse across steps, with byte-stable prompt prefixes reusing cached computation; parallel read-only batches can overlap I/O, executing in batches after a single inference.
- Tools and integrations: Connects to external tools via the Model Context Protocol (MCP), and can integrate with 1000+ apps (via Composio); supports custom skills to teach it repeatable workflows, and can be driven remotely from Telegram.
- Memory: SQLite-backed memory storage.
- Task loop: The local model decides the next action, the browser, files, and shell can execute in parallel, and risky steps pause for approval; the loop reuses cached context to maintain speed until a reply or the user stops it.
How to use:
- Terminal installation: A single curl command installs the CLI (Developer Preview v0.6.8), running on macOS (Apple Silicon), Linux x64 and arm64, Windows x64. Intel Mac is currently not built.
- Desktop app: Available from the download page, supporting macOS 14 or later (Apple Silicon), Windows x64 and ARM64, Linux x64 and arm64.
- Can also run via cloud models: Teams can use Atomic Agent with cloud models via OpenRouter, but cloud providers are used only when the user explicitly configures them.
Applicable scenarios:
- Automation tasks that need to run locally, privately, and offline, such as file organization (scanning, renaming, deduplication), document processing, code repository operations (cloning, reading issues, fixing, submitting PRs), browsing multiple sources and extracting citations, scheduled tasks (Cron monitoring Bitcoin prices, sports scores, and pushing to Telegram), etc.
- Privacy- or compliance-sensitive workflows, because by default no data is sent to the cloud.
Limitations and notes:
- The CLI is a developer preview (v0.6.8), and Intel Mac is currently not supported.
- Runs locally by default; cloud providers require the user to actively configure them.
- Free and open source, MIT License, no token bills; however, running it requires local hardware support (such as an everyday GPU).
- Atomic Agent is an NVIDIA Inception member (accepted in August 2026); the program does not involve equity, funding, or investment, and only provides developer resources, NIM API endpoints, and cloud credits; the project remains MIT open source and local-first, and does not require an NVIDIA GPU.
Performance reference: In the GAIA Level 1 benchmark (53 real tasks, same model, same hardware compared with Hermes), Atomic Agent scored 69.8% (37/53), while Hermes scored 58.5%, 11.3 percentage points higher; the average time per task was about 217 seconds, compared with about 351 seconds for Hermes, about 1.6 times faster.
