v3.0.0

Mirumoji 3.0.0 Refactor Overview

This is a structural and packaging rewrite of Mirumoji

• Backend + CLI are merged into a single, pip-installable mirumoji package, and the release, docs, and dev-container tooling are rebuilt around it

• The core immersion workflow is unchanged from 2.6.0

• The Launcher (CLI + Desktop GUI) is substantially expanded, and LLM support is no longer limited to OpenAI

Versioning Policy

• This is the baseline for the project's versioning policy

• Strict SemVer + Maintained Changelog Start Here

• There's Intentionally no 2.6.0 → 3.0.0 Changelog, since nearly everything moved internally, so this PR message answers 3 relevant questions instead, What Carried Over ? + What's New + How To Run It

What Carried Over From 2.6.0

The Immersion Workflow Is Unchanged

• Upload local videos, anime episodes, or audio for clickable tokenized Japanese subtitles with dictionary lookups

• Transcribe audio / generate subtitles with Whisper

• Get word / sentence breakdowns from LLMs, or prompt the LLM to refine the whisper-generated subtitles

• Save clips and export them to an Anki deck

• Organize your data (clips, LLM templates, files, transcriptions, ...) on the server by profile

• Self-host the Docker Compose Application with CPU / Local-NVIDIA-GPU / Modal Cloud-GPU Offload backend options

• Access the application via HTTPS from any device on your local network using the automatically generated self-signed certificate

What's New / Expanded in 3.0.0

Multiple LLM Providers (New)

2.6.0 Required an OpenAI API Key. 3.0.0 makes LLM features completely optional and adds Anthropic (Claude) + Google (Gemini) + Any Custom OpenAI-Compatible Endpoint support via a provider / model picker

CLI Launcher (Exapanded)

The 2.6.0 CLI had 5 commands (launch / shutdown / launch_local / build / gui ) driven by interactive prompts and a hand-managed .env. 3.0.0 rebuilds it on Typer / Rich + adds the status / logs / doctor / server / render commands + a managed-config surface (config set/delete/import/show/path/clear)

Desktop Launcher (Expanded)

The 2.6.0 flaskwebgui / PyInstaller window is rebuilt on Flet and gains a Settings panel where you can configure the transcription backend, image source, and LLM/Modal keys). It also has full environment checks, live status display, and docker compose log filtering

Modal Offload (Hardened)

Modal GPU jobs stream their media through a per-job ephemeral Modal Volume
instead of a baked image mount, so long media (multi-hour, multi-GB) transcodes
and transcribes reliably. Large uploads also stream at full speed rather than
being throttled at the reverse proxy

How To Run It

The New Setup Section contains detailed information on all of the ways that you can get Mirumoji running

Upgrading From 2.6.0

Warning

Your data does not carry over. The database schema changed in 3.0.0, so existing 2.6.0 profiles, clips, transcripts, and templates are not migrated. Treat 3.0.0 as a fresh install

Additional Details → Changed Surfaces

Package

apps/backend + apps/cli merged into one apps/mirumoji/ package published to PyPI as mirumoji

CI / CD

• 12 workflows are redesigned as an orchestrated release.yaml calling reusable _version / _images / _pypi / _pages / _desktop workflows

• Images are published to Docker Hub only (GHCR dropped)

Docs

• MkDocs Material custom Sumi & Shu css theme

• mkdocstrings-python API + TypeDoc frontend API

• awesome-nav structure

Dev Containers

• fixed builds + postCreateCommand bootstrap
• Changed from flake8 to ruff

Community

• Community files moved to .github/

• YAML issue forms

• Quality-gate PR template