Magnitude - Local AI Browser Agent with Built-in Models

TL;DR

TL;DR: Magnitude is an open-source AI agent with local models built in — no Ollama, no API keys, no cloud. Installs via npm and runs fully offline on macOS, Linux, or Windows WSL.

Source and Accuracy Notes

This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: magnitudedev.github.io — verified landing page for the open-source project
• Source repository: github.com/magnitudedev/magnitude — README read in full before writing
• License: Apache 2.0 — verified from LICENSE file in repo
• HN launch thread: news.ycombinator.com/item?id=43910133 — Show HN thread for Magnitude (145 points, verified)

What Is Magnitude?

Magnitude describes itself as “an open source agent with local models built in.” Unlike Ollama (which runs models) or Hermes (which orchestrates agents that use models), Magnitude combines both: it profiles your hardware, downloads compatible GGUF models, and runs them inside the agent process. Nothing else to configure.

From the README:

“Fully private and offline. Everything stays on your machine, including the models.”

The key differentiator is works out of the box: no separate model server, no OpenAI-compatible endpoint, no API key. You run npm install -g @magnitudedev/cli, cd your-project, and magnitude. The agent handles model selection and lifecycle.

Setup Workflow

Step 1: Install the CLI

npm install -g @magnitudedev/cli

Requires Node.js 18 or later. The package is published on npm as @magnitudedev/cli.

Step 2: Initialize a project

cd your-project
magnitude

On first run, Magnitude profiles your hardware (CPU, RAM, GPU if present) and downloads a suitable GGUF model from its catalog. The catalog includes models compatible with llm.cpp and similar backends.

Step 3: Add browser automation skills

Magnitude uses a skill system for extending capabilities. Browser automation comes from the agent-browser skill:

npx skills add vercel-labs/agent-browser

This drives a logged-in Chrome browser. Other available skills include xlsx, pptx, docx, and pdf for document work. Skills are sourced from skills.sh, a directory run by Vercel.

Step 4: Run the agent

magnitude

The agent starts a REPL-style session. You can issue natural-language prompts, and the agent will use its available skills (including browser automation) to execute tasks. All inference runs locally — no data leaves your machine.

Deeper Analysis

What “built-in local models” actually means

Magnitude bundles model support directly rather than requiring a separate inference server. It downloads GGUF-format quantized models and runs them via llm.cpp. This is architecturally similar to Ollama’s model management, but integrated into the agent loop rather than exposed as a standalone API server.

The README states:

“Profiles your hardware and recommends the best models for your machine.”

For a MacBook with Apple Silicon, it would select ARM64-optimized builds. For a Linux machine with an NVIDIA GPU, it would prefer CUDA-enabled variants. The exact model selection logic lives in the agent’s startup profiler.

Skills system

Skills are reusable capability modules. They are not bundled by default — you opt in via npx skills add. This keeps the core install lightweight while allowing browser automation, document processing, and similar extended features.

The skill registry at skills.sh aggregates community-contributed skill packages. The agent-browser skill from vercel-labs is the canonical example for web automation tasks.

Offline and privacy guarantees

Because models run locally and the agent is self-contained, prompts and files never reach a remote server. The README confirms:

“No token costs, API keys, subscriptions, or rate limits.”

This is a meaningful distinction from cloud-based agents that meter usage or route data through third-party inference APIs.

Open-source license

Magnitude is Apache 2.0 licensed. The full license text is in the repository at LICENSE. Apache 2.0 permits commercial use, modification, and redistribution with attribution — the same license used by many llm.cpp-family projects.

Practical Evaluation Checklist

• Installs via npm — npm install -g @magnitudedev/cli verified from npm registry
• No API key required — confirmed from README “No token costs, API keys”
• Runs fully offline — confirmed: “once downloaded, use without internet”
• Browser automation via skill — npx skills add vercel-labs/agent-browser from README
• GGUF model catalog — downloads compatible models; can also use Hugging Face models outside catalog
• OpenAI-compatible endpoint support — can connect custom inference endpoints per docs
• macOS, Linux, Windows WSL — platform support from README
• Apache 2.0 license — confirmed from LICENSE file

Security Notes

• Data never leaves the machine — local inference means prompts and context stay on-disk
• No telemetry or phone-home — not mentioned in README; the open-source nature allows self-hosting the skill registry
• Skill registry trust — skills from third parties (including Vercel) execute code in your local environment; review skill source before adding
• GGUF model sources — models downloaded from Hugging Face or Magnitude’s catalog; verify model provenance for sensitive workloads

FAQ

Q: How is Magnitude different from Ollama? A: Ollama is a model runtime — it downloads and serves models via an API. Magnitude builds on this by embedding the model runtime inside an agent framework, so the model lifecycle is managed alongside agent capabilities rather than as a separate service.

Q: What hardware do I need? A: The README states there is no fixed minimum — Magnitude profiles your hardware and selects accordingly. More RAM enables larger models.

Q: Can I use my own model? A: Yes. You can download compatible GGUF models from Hugging Face outside the catalog, or connect an OpenAI-compatible inference endpoint.

Q: Does it work on Windows? A: Yes, through WSL (Windows Subsystem for Linux). Native Windows is not listed as supported.

Q: Is the agent stateful across sessions? A: The default mode is REPL-based (per-session). The skills system can be extended for persistence, but the core agent does not ship with a built-in memory store.

Conclusion

Magnitude solves the “I want a local AI agent but I don’t want to set up Ollama plus an agent framework plus figure out the integration” problem. The model runtime is bundled, the install is one npm command, and browser automation is a skill away.

For developers who want a self-contained local AI agent without managing separate infrastructure, this is among the lowest-friction options available. The Apache 2.0 license means you can inspect, modify, and redistribute it freely.

If you need fine-grained control over which model or inference backend to use, the OpenAI-compatible endpoint support provides that escape hatch without abandoning the local-first model.

Source last checked: 2026-08-23 (commit a3f09a0, 1494 stars on GitHub)