A local-first AI agent developed by Atomic Bot. Released as open source under the MIT license, it runs AI models on your own machine through llama.cpp and takes on work such as browser operation, file editing, shell command execution, and git inspection. The agent’s control loop, memory, and traces are all stored locally as plain files and SQLite databases, and nothing is sent to the cloud by default. No API key and no usage-based billing are required, so it costs nothing to operate as long as your machine can run the model. Released in April 2026, it supports macOS (Apple Silicon), Linux, and Windows.
Key Features
- Local execution at zero cost: Because models run on your own hardware with llama.cpp as the engine, there is no per-token bill. The default model is a small, lightweight one that fits within the reach of an ordinary laptop
- Grammar constraints keep tool calls valid: GBNF (llama.cpp’s grammar notation) forces the model’s output format, so even small local models produce structurally valid tool calls. The classic local-model failure of “broken JSON halts the run” is designed out rather than patched over
- Parallel browser, file, and shell operation: It handles browser automation via Playwright (navigate, click, type, read), file reads and writes, approval-gated shell commands, web search, extraction from PDF and Office documents, and git inspection. Multiple tools can run concurrently
- SQLite-based memory: Facts, notes, and learned procedures live outside the prompt and are injected into context only when needed, so working assumptions carry across sessions
- Extensible through MCP: Register Model Context Protocol servers in the config file and their tools join the agent’s toolbox. Untrusted servers are approval-gated by default
- Cloud options remain open: When you need them, OpenAI-compatible servers, OpenAI, OpenRouter, and flat-rate CLIs such as Claude Code and Codex can serve as the inference provider, so you are not locked into local-only operation
Pricing
| Plan | Price | Key features |
|---|---|---|
| Open source | $0 | All features available. MIT license, commercial use permitted |
The software itself is free, and no additional cost arises as long as you run local models. Provider fees apply only if you choose an external cloud model as the inference source.
Pricing is current as of August 2026. Check the official site for the latest information.
Pros & Cons
✅ Pros
- Your data never leaves your machine, which suits situations where work content, files, and history should not be handed to an outside service
- You can start without issuing an API key or registering a credit card, and heavy use does not increase any bill
- Open source under the MIT license, so the implementation can be read, modified, and used commercially
- Grammar constraints make tool calls comparatively stable even with small models
- MCP support lets you bring existing tool assets along as they are
⚠️ Cons
- As of August 2026 it is in developer preview (the v0.2 line), so specification changes and bugs are possible
- Memory and disk are needed to run a model, and local inference performance depends directly on your hardware
- On macOS only Apple Silicon is supported; no Intel Mac build is provided
- Installation assumes the command line, so it lacks the ease of a GUI application
- The reasoning quality of small local models falls short of major cloud models in some situations
Comparison with Similar Services
| Item | Atomic Agent | Goose | Open Interpreter | Claude Code |
|---|---|---|---|---|
| Provider | Atomic Bot | Block | Open Interpreter | Anthropic |
| License | Open source (MIT) | Open source | Open source | Commercial (subscription) |
| Where models run | Mainly local (cloud optional) | Mainly external LLMs | Mainly external LLMs | Cloud (Claude) |
| Main targets | Browser, files, shell, git | Files, shell, extensions | Code execution, files | Codebase, shell |
| Extra cost | None when run locally | Model usage fees | Model usage fees | Subscription fee |
Who Is It For
- People who want to try an AI agent in an environment where business data and internal files must not be sent outside
- People who want to run an agent for long stretches and at high frequency without worrying about usage-based billing
- People exploring practical use of local LLMs such as the Qwen, Gemma, and Llama families
- Developers who want to read MIT-licensed code and adapt it to their own workflow
- People already running MCP servers who want to add a fully local execution environment
Summary
Atomic Agent is an open-source project that tries to make “an agent that works usefully even on local models” real through concrete design choices: GBNF grammar constraints, parallel tool execution, and SQLite-backed memory. Freedom from the cloud means your machine’s performance becomes the perceived speed, and its developer-preview status brings rough edges. Even so, the combination of zero cost and fully local operation is hard to replace. For work with strict privacy requirements, or for running an agent without watching the meter, it is worth trying first.