Questions I get asked

A free, open-source, local-first AI session chronicler for D&D and other tabletop RPGs: it transcribes your session recording offline with Whisper (optional add-on), then runs a 6-phase LLM pipeline to produce a polished narrative recap grounded in your campaign's own glossary and lore.

Yes. The software itself is MIT-licensed and free forever — no subscription, no account, no telemetry, no paywall. The only money you might spend is on a cloud LLM API key — Gemini and OpenRouter are both pay-as-you-go, and Tomes is engineered to keep that bill small (prefix caching, per-tier chunk sizing, guided routing recipes). A three-hour session lands somewhere around $1–$5 depending on the routing you pick; what it costs has the current figure for each one, priced against the live catalogue. If you already pay for Claude or ChatGPT, the Claude Code and Codex CLIs spend that allowance instead of API credit. You can also run fully offline with Ollama, LM Studio, or Unsloth — in which case it costs nothing to run. See pricing.md.

Yes — install the Claude Code add-on. It runs the pipeline through the claude CLI you've already logged into, billed against your plan with no API key. One caveat: subscriptions meter headless use in rolling ~5-hour windows, and the pipeline is token-heavy — in testing, one full session used up to ~60% of the allowance in a 5-hour window. Budget for roughly one or two sessions per window, and consider letting Claude Code do the grounding + chronicle while Chronicle Reforge handles the cheaper extras + recap on Gemini.

No. Chronicle Reforge (Tome of Lore → Saved Chronicles) re-runs just the later phases of a finished chronicle on a different model — keep the chronicle and redo the quotes / jests / gore and condensed recap on, say, Gemini, or regenerate the chronicle itself in a persona voice. It saves a new entry, so your original stays put. And if a provider declined specific chunks, those are flagged in the output and a one-click Review & Repair Refusals panel re-processes only those on another provider.

No. Audio transcription happens locally, on your CPU or GPU, using faster-whisper (via the Audio Transcription add-on). The raw audio never touches the network. Only the resulting text transcript + prompt scaffolding is sent to the LLM provider you've configured — and only when you click Run. If you install the Local LLMs add-on and route through Ollama / LM Studio / Unsloth, even the text stays on your machine. See privacy.md for the full trust model.

No, but it helps a lot — and it only matters if you're using the Audio Transcription add-on. Worth being precise, though: it has to be an NVIDIA card. Transcription runs on faster-whisper, whose backend (CTranslate2) implements only two devices, cuda and cpu. There's no ROCm, Metal or Intel Arc backend, so an AMD, Intel or Apple GPU isn't used at all — you get processor speed, same as a machine with no graphics card. That's the engine's limitation, not a preference. With an NVIDIA card, a 4-hour session drops from roughly 90 minutes on CPU to about 12. When installing it (Settings → Transcription), it auto-detects your GPU via nvidia-smi and installs the matching torch wheel; if no GPU is present, it falls back to CPU automatically.

Short answer: Google Gemini (paid) for the cheapest cloud option overall, OpenRouter if you want the widest choice from one key (around 400 models, including every Anthropic and OpenAI model at pass-through rates), a subscription you already pay for via the Claude Code or Codex CLI, or a local model via Ollama if you want to spend nothing and keep everything offline. You can mix and match per pipeline phase in Settings → Providers & models, where each phase gets one row. For most groups, a cheap model on the mechanical phases plus a strong one on the prose phase is the sweet spot. See providers.md.

Yes. Pathfinder, Call of Cthulhu, Blades in the Dark, Vampire: the Masquerade, Daggerheart, Fabula Ultima, homebrew systems — anything you'd record as a Discord session or upload to YouTube. The chronicler doesn't know or care about the system; it grounds itself in your glossary and your lore PDFs, which is where system-specific terminology lives.

Two different answers, because they're two different categories.

Otter, Descript and NotebookLM have no concept of "this is a TTRPG session" — they'll mis-transcribe every fantasy proper noun and they upload your audio to their cloud. NotebookLM is closest in spirit, but it's a generic research tool: you'd upload a transcript by hand, prompt it by hand, and get generic prose rather than a chronicle.

The services built for tabletop — Tabletop Scribe, DnD Scrybe, Kazkar, SessionKeeper, Saga20 and others — are a fairer comparison, and several are good. They ground names against your lore too, so that isn't the difference. Three things are. Your audio never leaves your machine, because Whisper runs locally rather than on someone's server. Attribution is known rather than inferred, because Craig hands over one track per player instead of one mixed recording. And you choose the model, the narrator and the prompts, because it's MIT-licensed.

The cost of the middle one is real: they work from a browser and this doesn't. See comparison.md for the full breakdown, including when to pick one of them instead.

No — Craig is one workflow, not a requirement. The lowest-friction path is the YouTube .sbv workflow (works without any add-on): if you already upload sessions to YouTube (public, private, or unlisted), download the auto-caption file from YouTube Studio and drop it in. Craig gives you per-speaker attribution throughout the chronicle, which is great if you want it, but it's optional. See workflows.md.

An optional capability the core app doesn't need. The core program does paste-a-transcript chronicling with cloud LLMs on its own.

There are seven, and the useful distinction between them is whether turning one on puts bytes on your disk. Audio Transcription does — it downloads a Python environment — so it has an Install button, a live log and a restart. The other six (whisper.cpp bridge, Local LLMs, Chronicle Personas, Claude Code, Codex, Obsidian Vault lore) ship with the app and are always available; whether one is usable is a detection question, not an installation question.

See add-ons/README.md.

Almost. You need to install Node.js (GUI installer from nodejs.org), and you need to either clone the repo or download the ZIP. After that, it's double-click setup.bat (Windows) and then double-click Start_Tusks_Tomes.bat. Python is only required if you want audio transcription — and even then, you can skip it on first install and add it from Settings → Transcription later. The full beginner walkthrough is in beginner-guide.md. A signed Windows installer that bundles everything is on the roadmap.

Roughly 15–20 minutes end-to-end with a GPU: ~12 minutes for Whisper transcription (with the Audio Transcription add-on), ~3–5 minutes for the LLM pipeline (Gemini Flash is faster than Claude Sonnet here). On CPU, the transcription step balloons to ~90 minutes but the LLM step is unchanged. The dashboard shows live progress and ETAs.

No. Your chronicles (Sessions/<campaign>/...), encrypted keystore, glossary, speaker mappings, routing, personas, and uploaded session audio all live outside the repo in your platform-standard config and data directories. The in-app updater (Settings → Maintenance) does a guarded git pull + npm install and refuses to clobber uncommitted edits. See configuration.md for exactly where things live on disk.

You're covered. When the LLM provider returns a 429 (quota exhausted), Tomes opens a four-option dialog mid-pipeline:

• Stop and export what we have — aborts the run; downloads whatever the pipeline produced so far as Markdown.
• Slow down (3× longer between calls) — keeps the same key but paces calls more conservatively for the rest of the run. Disabled if the quota looks daily (waiting longer doesn't help an empty bucket).
• Pause and save for later — writes a full on-disk checkpoint. Close the app, come back tomorrow when the daily quota refills, click Resume on the banner above the Chronicle card. The pipeline picks up at the exact chunk it stopped on.
• Switch keys for the rest — only available when you've configured a fallback key. Most commonly used in Smart Budget mode for Phase 4 extras: if your optional free-tier Gemini key stalls, this jumps the rest of Phase 4 to your paid key.

Daily quotas typically reset at midnight UTC; per-minute limits clear in 60 seconds. None of your work is lost regardless of which option you pick.

Three levers, in roughly the order most users reach for them:

1. Pick a cheaper rung in guided routing. Settings → Providers & models offers a ladder of complete per-phase recipes, from maximum quality down to free. Moving off the flagship-everywhere rung cuts the cost several times over. Quality on the long-form chronicle phase is noticeably weaker on fast-tier models — A/B-test a 30-minute slice before committing if quality matters.
2. Use Smart Budget instead. Same UI; a different preset. Paid Flash for grounding + audit, Paid Pro for the chronicle (the quality phase), Paid Flash-Lite for condense. If you've also configured an optional free-tier Gemini key, Smart Budget routes Phase 4 extras to it (the only phase that ever uses your free key). Roughly half the cost of Pro-everywhere, with the chronicle quality preserved. What it costs has the current figure.
3. Mix-and-match per phase manually. Same panel, one row per phase. Use the cheap model on Phase 1 (Grounding) and Phase 4 (Extras) where prose quality barely matters, and keep the strong model on Phase 3 (Chronicle) where it does.

See providers.md for honest per-provider cost estimates and the architectural cost-reduction work shipping in the current release.

** If you also have a free-tier Gemini key, the better home for it is Tusk's Vault** — the upcoming AI-chatbot companion that lets you ask questions of your campaign lore. Vault is a single retrieval-augmented query per turn, far lower token use than Tomes' six-phase pipeline, so a free quota carries Vault's workload comfortably. Vault is due to release soon.

Not yet — it's a roadmap item. A shared community Discord for Tusk's Tomes and the companion project Tusk's Vault (#support, #feature-ideas, #showcase, #dev-talk) will be spun up once there's enough signalled interest.

In the meantime, the feedback form is where to drop setup questions, feature ideas, chronicles you want to share, and — most importantly — a "yes" to the "would you join a community Discord?" question. That signal is what triggers the launch.

Yes. Tusk's Tomes is MIT-licensed — you can use it personally, commercially, modify it, ship it inside your own product, sell support around it, anything. The only requirement is that the MIT copyright notice travels with the code. We'd love to hear about commercial uses via the feedback form, but you don't owe us anything.

Open an issue (there are guided templates for bug reports and feature requests), or send a PR straight from your fork. For larger changes, sketch the idea via the feedback form first (a community Discord with a #dev-talk channel is on the roadmap). The full contributor guide is CONTRIBUTING.md, and the community standards are in the Code of Conduct.