Great Lakes Institute for Applied Intelligence
Back to Great Lakes

Build

Small tools and practical experiments that turn what we’re learning into something useful.

Featured project

Built by the Institute, free to use and free to inspect.

gumee

A simpler place to experiment with agents.

gumee is a free, open-source desktop app for people who want to explore AI agents without wrestling with developer infrastructure. You don’t need Node, Python, Docker, Git, a local model, or dev servers. Attach a few files, describe a task in plain language, and watch a real agent plan, use specialized tools, and generate outputs inside a safe, bounded workspace.

  • Version0.2.1
  • LicenseMIT
  • Runs onMac, PC, Linux
  • PriceFree

  • No developer overhead

    A packaged desktop application. Download it, launch it, paste an OpenRouter key, and start experimenting.

  • Visible tool steps

    Built on LangChain’s TypeScript Deep Agents runtime, with real-time streaming and durable SQLite checkpoints.

  • Bring your own key

    gumee is free. Models run through OpenRouter on your own API key, held in your OS keychain or for the session only.

  • Bounded workspace

    Everything runs on your machine. Tools read only the inputs you attach and write only to dedicated run folders.

Download gumee

Latest release, version 0.2.1. Read text from images and scanned PDFs with built-in English OCR. Open source under the MIT license.

This matches the computer you’re on.

gumee for Mac

Apple Silicon (M1, M2, M3, M4), macOS 12 or later. Unsigned installer.

gumee-0.2.1-macos-arm64-unsigned.dmg Download for Mac

All releases Documentation

Install it

  1. Open the downloaded .dmg file.
  2. Drag gumee into your Applications folder.
  3. Launch gumee from Applications or Spotlight.

gumee is built without paid Apple developer certificates, so macOS may block the first launch. Right-click gumee in Applications and choose Open, or allow it under System Settings › Privacy & Security.

Verify the download shasum -a 256 gumee-0.2.1-macos-arm64-unsigned.dmg

This matches the computer you’re on.

gumee for PC

Windows 10 or 11, 64-bit. Installer.

gumee-0.2.1-win-x64.exe Download for PC

All releases Documentation

Install it

  1. Run the downloaded gumee-0.2.1-win-x64.exe installer.
  2. Follow the steps. It installs for your user and adds a Start Menu entry.
  3. Launch gumee from the Start Menu or your desktop.

If Windows SmartScreen says “Windows protected your PC”, choose More info, then Run anyway to continue.

Verify the download certutil -hashfile gumee-0.2.1-win-x64.exe SHA256

Linux, or building from source

On Linux, grab the 64-bit gumee-0.2.1-linux-x86_64.AppImage from GitHub Releases.

To inspect the code or build it yourself, clone the repository and run npm ci && npm run dev in the gumee/ directory. The source is on GitHub.

How to experiment with gumee

Three steps, in order, once the app is installed.

  1. Connect your OpenRouter key

    gumee uses your own model provider account, so you never pay subscription markups or hidden fees. Get a key at openrouter.ai, paste it into gumee under Settings, and choose Test.

    Two agent profiles come built in:

    • Erie is fast and economical, for everyday tasks, quick summaries, and drafting.
    • Huron reasons harder and holds more detail, for multi-page document synthesis and structured calculations.

  2. Pick a sample dataset

    You don’t need to prepare any files. Three starter exercises ship with the app, on the home screen:

    • Compare two documents reads two revisions of a proposal line by line and reports what changed and what went missing.
    • Meeting notes to action items turns an unstructured transcript into a list with owners and priorities.
    • CSV data analysis inspects a table, computes statistics in code, and writes up what it found.

  3. Run it, then check its work

    Choose Run and watch. Unlike an opaque chat window, gumee streams each step as it happens:

    • Inspect the tool calls. Expand any step to see which tool the agent chose, which pages of a PDF it read, and the exact computations it ran.
    • Preview the output. Open generated markdown, CSVs, or PDFs inside the app, or reveal them in a folder.
    • Try a variation. Change the prompt or switch models on a finished run, then compare the two outputs side by side.
Waves breaking over a rocky Lake Superior breakwater, with a lighthouse standing at its far end.
Something built, out where the weather is. Lake Superior photo collection

How we approach a build

  • Keep the problem small

    Choose one useful job and understand what a good result would look like.

  • Try it in practice

    Make something simple, work through real examples, and find out where it needs attention.

  • Share what we learn

    Explain what works, what doesn’t, and what someone should know before using it.