Microsoft's agent that runs on your laptop, not their cloud
An experimental browser-and-filesystem agent built to work with small models instead of GPT-4-scale compute.

What it does
MagenticLite is a local agentic application that automates tasks across web browsers and your file system — think filling expense forms, scraping recipe prices, booking restaurants, or organizing folders. It runs as a local Python service you access through a web UI at localhost:8081.
The interesting bit
The architecture is deliberately split: a lightweight “orchestrator” model (MagenticBrain) handles planning and user interaction, while a specialized browser-use model (Fara) does the actual web manipulation. This lets the system run on modest hardware instead of requiring frontier-scale models. Browser sessions are sandboxed inside a lightweight VM called Quicksand, so the agent can’t wander your filesystem without explicit permission.
Key highlights
- Designed for small models; explicitly avoids dependency on frontier-scale compute
- Human-in-the-loop by default: stops for approval before critical actions
- Browser sandboxing via Quicksand VM; local file access is gated
- Ships on PyPI (
magentic_ui>=0.2.0); runs on macOS and Windows WSL - Previous 0.1 branch (frontier-model-optimized) still available
Caveats
- Requires bringing your own model endpoint; no built-in model hosting
- macOS and WSL only for quick-start; other platforms need manual setup
- Has a documented limitations file — the maintainers are upfront about what it can’t do
Verdict
Worth a look if you want a local, approval-gated agent that doesn’t phone home to OpenAI. Skip it if you need turnkey model hosting or a fully autonomous agent that never asks for permission.