MagenticLite (Magentic-UI) microsoft/magentic-ui

Microsoft Research agent app for web research, form filling and file tasks, with a sandboxed browser VM and approval before critical actions.

Creator kit + Compare GitHub ↗ Install guide (official) ↗

MagenticLite (Magentic-UI) at a glance

Checked 16 Oct 2026

MagenticLite (10.1k stars, MIT, v0.2.1 May 2026) is the new version of Microsoft's Magentic-UI. A local web app pairs an orchestrator model (MagenticBrain) with a browser-use model (Fara1.5-9B), both MIT-licensed weights. The browser runs in a lightweight VM sandbox, and the agent checks in before critical actions; you can steer or take over. You must host both models behind OpenAI-compatible endpoints (about 48 GB and 96 GB GPU, or paid Hugging Face or Azure endpoints). Tested on Apple Silicon and Windows via WSL2. An experimental research app: no slides, schedules or app connectors.

Cite this pageRepoQuarry. “MagenticLite (Magentic-UI).” Checked 16 Oct 2026. https://repoquarry.dev/projects/magentic-ui

How much of Manus does it replace?

2 of 6 jobs. Sandboxed browsing plus local file work with human approval; needs large self-hosted models; no slides, schedules or connectors.

Autonomous web browsing Partial replacement

Deep research reports Not covered

File and code tasks in a sandbox Partial replacement

Slides generated from a prompt Not covered

Scheduled agent tasks Not covered

Connectors to your apps Not covered

For deep research reports, slides generated from a prompt, scheduled agent tasks, connectors to your apps, see other Manus alternatives. You may also need to build an interface, an API server and hosting yourself.

Facts, with sources

Every answer links to where we found it. We read the documentation; we did not install it.

Report a correction

QuestionAnswerSourceChecked

Runs on my machine? Yes, local web app; models run on endpoints you host or rent Works across the browser and your local file system. github.com · magentic-ui#readme ↗ today

Code licence MIT Microsoft, and any contributors, grant you a license to any code in the repository under the MIT License. github.com · magentic-ui#license ↗ today

Model weights licence MIT for both MagenticBrain and Fara1.5-9B weights huggingface.co · Fara1.5-9B ↗ today

Commercial use Yes, MIT code and MIT model weights github.com · magentic-ui#license ↗ today

Paid services needed? Nothing paid if you host both models on your own GPUs with vLLM; otherwise paid Hugging Face or Azure Foundry endpoints Hugging Face Inference Endpoints. Managed GPU hosting, billed per minute. github.com · model-hosting-guide.md ↗ today

Hardware Self-hosted models need GPUs with about 48 GB (Fara) and 96 GB (MagenticBrain); the app needs KVM on WSL for the sandbox VM plan for at least 48 GB of GPU memory for Fara and 96 GB for MagenticBrain. github.com · model-hosting-guide.md ↗ today

Operating systems macOS on Apple Silicon and Windows x64 via WSL2 tested; native Linux untested; Windows ARM64 unsupported It covers macOS and Windows (WSL); Linux is similar to the Windows/WSL path. github.com · installation.md ↗ today

Setup effort Medium for the app (uv pip install); hard overall because you must host two model endpoints uv pip install "magentic_ui>=0.2.0" github.com · installation.md ↗ today