Atomic Agent
Open source agent (MIT licence) that runs in your terminal, with the control loop and all state on your machine: sessions, memory, traces and settings are local files. It reads and edits your files, runs commands once you approve, browses the web and calls tools over MCP. It runs open-weight models on your hardware through llama.cpp, which it downloads and keeps current for you, or a remote model on your own key. Its Fusion mode mixes both: one model plans and hands the bulk of the job (wide reads, first drafts, tests) to a pool of workers, then checks and merges their answers. The usual setup puts a remote model on the plan and local models on execution.
Strengths
- Open source (MIT), state and traces kept as local files: you can read back every prompt and every tool call
- Fusion mode built in: one model plans, workers execute in parallel (2 by default, adjustable)
- Local models managed for you: the tool downloads the llama.cpp backend and the models, and picks the GPU backend on its own
- Dangerous commands ask every time, even after you granted commands for the whole session
Limitations
- Young project: version 0.7.0 released on 9 November 2026, little hindsight on how it holds up in real use
- With local workers, how many tasks run side by side is capped by your machine
- The repository says so itself: secret redaction and environment filtering are not complete isolation
Best for
- Developers with a decent GPU who want to cut their calls to a paid model
- Teams that want model calls to stay on their own machine