Plume

Why I built it

A lot of desktop AI apps are built with Electron and use a lot of memory before the model has even done anything. That is a problem on laptops with limited RAM, especially when a local model needs the same resources.

I built Plume so the workspace around the model stays lightweight. It uses Tauri, Rust, and the macOS system WebView, leaving more of the machine for local inference and the user's actual work.

What it does

Plume is a local AI workspace for Mac. You can chat with Apple On-Device, a Plume-managed Qwen Coder model, or Ollama. Chats and branches are saved locally.

You choose exactly what goes into the conversation: project files, Browser text, screenshots, memories, or Library items. Plume can use attached Browser text to make a research note with source links, and it can export the result as Markdown when you ask. The Browser stays under your control.

For code changes, the model can propose a diff, but Plume validates it first and only writes it after you press Apply. Revert uses the saved checkpoint. There is no broad shell access or hidden tool runner.

How I built it

The desktop shell is Tauri 2 and Rust. The interface uses React, TypeScript, and CodeMirror. Local model support is built around Apple Foundation Models, MLX-LM, MLX-VLM, and Ollama compatibility. There are no default cloud calls.

How I used Codex and GPT-5.6

I used Codex with GPT-5.6 Sol throughout Build Week. It helped me break work into smaller pieces, implement and test the Rust and React parts, review the trust and redaction boundaries, run the full verification gates, and test the packaged Mac app. I still made the product decisions and checked the actual code and tests before accepting changes.

What was difficult

Getting a model response was not the hard part. The hard part was keeping context exact, keeping local and project data separate, and making sure the model could not quietly read or change more than the user approved. A few bigger agent features are still future work, so the submission only claims what is working now.

Built With

• react
• rust
• tauri
• typescript