Ogmios

Practise English by talking with Claude. You pick a topic and a level, speak or type freely, and nobody corrects you mid-sentence. When you finish, you get one thing to focus on, two smaller notes, and better ways to say what you said. Ogmios remembers your recurring mistakes across conversations and steers the next ones toward them until they are gone.

• Feedback at the end, and only what matters. One focus, two minor points. The rest waits in memory.
• You fix it first. For grammar, you see your own sentence and try to correct it before the answer appears.
• Memory of your mistakes. Each one moves through detected → in focus → improving → mastered. It counts as mastered only when you get it right in real conversations, not just in drills.
• Short drills when you want them. Two minutes, mixing the structure you are working on with the others.
• Voice or text. Speech recognition runs on your machine. You can edit the transcript before sending.
• Local and private. Conversations, audio and progress stay in a SQLite file on your computer.

Install

macOS (Apple Silicon), with Homebrew:

brew tap ilien-dev/ogmios https://github.com/ilien-dev/ogmios
brew install --cask ogmios

Windows: download Ogmios_x64-setup.exe from the latest release and run it. The installer is not signed yet, so SmartScreen warns about it: choose More info → Run anyway.

After that Ogmios updates itself: it checks for a new release at start-up, and Settings → About Ogmios shows the version and installs the update.

Connecting to Claude

Ogmios needs one of these:

1. An Anthropic API key (recommended). It is stored in your system's keychain.
2. Your local Claude Code. Ogmios runs the claude you installed and logged into yourself. It never sees your credentials. Anthropic may limit subscription use in third-party apps, so this mode can stop working at any time.

A 20-minute conversation costs roughly $0.20 with Sonnet 5 on the API. That is an estimate, not a measurement.

Development

Requirements: Bun 1.3+, Rust stable, and the Tauri prerequisites for your platform (on Linux, also libasound2-dev).

bun install
bun run hooks:install   # formatting check before each commit
bun run tauri:dev       # the desktop app
bun run dev             # the interface alone, on a fake backend
bun run check           # types, lint, format, tests (TypeScript and Rust)

The speech recogniser for Linux x86_64 is in src-tauri/vendor/parakeet/. For another platform, run scripts/build-parakeet.sh (needs CMake, Ninja and a C++ compiler).

Layout: src/ is the React interface, src-tauri/ the Rust backend (data, error memory, speech), sidecar/ the process that talks to Claude, shared/ the types they exchange. The product spec is in docs/SPEC.md.

Licence

AGPL-3.0-only. See LICENSE. The speech recogniser (parakeet.cpp, ggml) is MIT; the Parakeet models are CC-BY-4.0 by NVIDIA.