Requirements

What	Version	Why	Check it with
Node.js	20 or newer (LTS)	The app is a Node program. This is the one hard prerequisite.	node -v
npm	comes with Node	Installs the libraries. You don't install this separately.	npm -v
Git	any recent	How you download the project, and how the in-app updater fetches new versions. Without it you'd be re-downloading a ZIP by hand every release.	git --version
A way to reach an AI model	—	Pick one of the three options below. The pipeline has nothing to run on otherwise.	—

The model — pick one of these three

You need exactly one. They're interchangeable and you can switch later.

Option	Cost	Best for
A paid API key — Gemini direct, or OpenRouter for around 400 models on one key	Roughly $1–$5 a session	Most people. Gemini is the best value for this particular workload.
A subscription you already pay for — Claude Code or the Codex CLI	$0 extra	Anyone already paying for Claude or ChatGPT. See Claude Code / Codex.
A local model — Ollama, LM Studio or Unsloth	$0	Total privacy, no network. Read the quality caveats first — below ~15–20B the prose gets noticeably worse.

You do NOT need Visual Studio Build Tools, whatever older guides say. Nothing in the dependency tree compiles native code.

None of this is needed to get a chronicle. All of it improves what you get.

What	Why you'd want it	Without it
Python 3.10–3.12	Runs Whisper, which transcribes your recordings locally. On Windows, tick "Add python.exe to PATH" during install.	You supply a transcript another way — pasted text or YouTube captions.
An NVIDIA GPU specifically	Transcribing a 3-hour session drops from several hours to 20–30 minutes. Detected automatically via nvidia-smi. Read the note below before assuming your card counts.	Whisper still works, on your processor, slowly. Or use the YouTube route.
Discord + Craig	Craig is a Discord bot that records one audio track per person, which is what makes speaker attribution reliable end to end. Craig's free tier is enough.	Single-stream audio, so the AI has to guess who said what — and it will ask you a lot of questions.

On GPUs — it really is NVIDIA or nothing, and that's worth spelling out

"I've got a dedicated graphics card, I'll be fine" is a completely reasonable assumption and an expensive one to get wrong, so here are the actual facts.

Transcription uses faster-whisper, which is built on CTranslate2. That library implements exactly two compute devices: cuda and cpu. There is no ROCm backend, no Metal backend, no Intel Arc backend, no DirectML backend.

So an AMD Radeon, an Intel Arc, or the GPU in an Apple Silicon Mac is not used at all. This isn't "slower but works" — the card sits idle and everything runs on your processor. A $900 AMD card gets you exactly the same transcription speed as no graphics card whatsoever.

That isn't a preference for NVIDIA on my part; it's the only GPU backend the engine has. And before you go looking for a Whisper build that does use your card — I did, and here's what's actually out there:

Project	Prebuilt GPU backends	Verdict
whisper.cpp	CPU, BLAS (still CPU), CUDA	Has a Vulkan backend in source only — it isn't in any release build. Compiling it yourself works.
sherpa-onnx	CPU, CUDA	No DirectML build shipped.
Const-me/Whisper	DirectCompute — any DirectX 11 GPU	Actually works on AMD and Intel. Windows-only, no release since July 2023.

So the honest summary: there is no maintained, ready-made Whisper build that will use an AMD or Intel GPU. Const-me/Whisper does, and has been unmaintained for years. Everything else prebuilt is CPU or CUDA.

Which leaves you three practical options:

1. The YouTube route — Google's hardware does the transcribing, no GPU of any kind needed. This is what I'd suggest.
2. CPU transcription. Slow, but it works if you set it going and walk away.
3. Transcribe elsewhere and paste it in — Workflow A accepts a transcript from any source, so a self-compiled whisper.cpp or Const-me output drops straight in.

Apple Silicon is at least the best case for the CPU path — those cores are quick enough that CPU-only transcription hurts far less than it does on an older x86 laptop.

Python 3.13+ is untested, not broken. The restriction originally existed because PyTorch shipped no 3.13 builds at all. Upstream has since released them and this project's requirements permit them — but nobody has verified the whole stack on 3.13 here, so setup warns rather than blocks. Try it if you like, and please report what happens. If it fails, install 3.12 alongside and the Whisper setup picks it up automatically.

Everything Python-related lives inside vendor/python-venv/, a self-contained environment. Uninstalling deletes that folder; your system Python is never touched.

Genuinely optional. Add them if they appeal.

What	What it gets you
Obsidian	Somewhere pleasant to keep campaign lore. The vault module reads your notes to get names right — though any folder of markdown works, so this is about your comfort, not a requirement.
Claude Code or the Codex CLI	If you already pay for Claude or ChatGPT, routes the mechanical phases through that allowance instead of API credit. Roughly 85% cheaper than the baseline.
Ollama / LM Studio	Run some or all phases offline. Good for the cheap mechanical phases even if you keep the chronicle itself on a cloud model.
A second, free-tier Gemini key	Used only for one small phase under the Smart Budget preset. Shaves a little off the bill.
Windows Terminal	The launcher opens in a tidy tab instead of a bare console window if it's installed. Purely cosmetic.
Tusk's Vault	The sibling project — ask questions of your finished chronicles in Discord. See vault.md.

What	Note
ffmpeg	Comes with the ffmpeg-static package. No separate install on any normal setup.
PDF and DOCX readers	Pure JavaScript, installed with everything else.
The Whisper model	Downloaded by the audio add-on when you install it, into vendor/python-venv/.

None of this is in the repo — it's all made at runtime.

Path	What	Created by
node_modules/	The JavaScript libraries (~360 MB)	npm install
.env	Per-machine settings	setup.bat / setup.sh
vendor/python-venv/	Whisper (1.5–2.5 GB)	The audio add-on, if you install it
Sessions/<campaign>/*.md	Your finished chronicles	The app, when you save one
%APPDATA%\tusks-tomes	Settings, glossary, encrypted API key	Saving anything in Settings
%LOCALAPPDATA%\tusks-tomes	Recordings, transcripts, run checkpoints	Uploading audio or pausing a run

A fresh clone contains no personal state at all — source, scripts and templates only. For the full account of what's written where and what could go wrong, see How safe is this?.