Free · Open source · Runs on your own machine
Point it at a recording of your D&D, Pathfinder or Call of Cthulhu game — or at a transcript you already have — and get back a narrative account of what actually happened. Spelled the way your table spells things, because it reads your campaign's own lore before it writes a word.
Install it Read what it writes
What the microphone heard
“Kaz eel swears on the crimson cathedral and draws the sun blaid”
What Tomes wrote down
“Kaziel swears on the Crimson Cathedral and draws the Sun Blade”
Why this exists
I wanted a record of my campaign. Not minutes — the actual story. The bit where the plan fell apart in an interesting way, the joke that had everyone crying, the line the DM delivered that made the table go quiet for a second.
The obvious tools didn't fit. They're built for meetings, and they have no idea who your characters are — they hear an invented fantasy name, guess a spelling, and then use that guess with total confidence for four hours straight.
And a session isn't minutes. What you want back is the story: something you would actually read again, the lines worth quoting, and a recap short enough that whoever missed the last one can catch up before they sit down. That is as much use to a player as it is to whoever runs the game.
So I wrote this instead.
It's a niche tool for a niche group of people, and I'm fine with that. If you're one of them, hello.
What comes back
A single run produces everything below. You choose which of them you actually want; nothing is mandatory except the chronicle itself.
One session — three or four hours of six people talking
A chapter of your campaign's book rather than a summary of it — continuous narrative prose, about 16,000 words for a full session.
One paragraph and catch-up bullets. What you send whoever missed last week, so they sit down already knowing.
Quotes, jests and the gorier highlights as their own lists. The stupid thing someone said in hour three, still there in a year.
Everything it could not resolve, asked rather than guessed at.
The raw text with every name corrected — and the end of “I definitely never agreed to that”.
.docx and Markdown, into a folder you choose.
Roused to their duties, the two clerics ascended to the top deck, their shared darkvision piercing the pitch-black harbour. To pass the hours, Grendal conjured a Leomund's Tiny Hut, projecting bright, magical cartoons across the interior of the dome. Ostyn added minor illusion hand puppets, turning the cold watch into a bizarre, domestic pillow-fort cinema.
That is one paragraph of one scene. Read a full session, the recap and the quotes →
How it compares
Often you should. Most people find this page having already tried a meeting transcriber — good software, built for standups, where nobody is called Vellichor the Pale, nobody needs the joke from hour two written down, and “the group discussed next steps” is a useful sentence.
There is also a set of tools built specifically for tabletop — Tabletop Scribe, DnD Scrybe, Kazkar, SessionKeeper, Saga20 and more arriving. They are made by people who play, several handle speaker attribution well, and they will take a recording without asking you to install anything. If one of them suits your table, use it.
These are the things worth checking first, including the one where this project comes off worse.
| Capability | Tusk’s Tomes | Tabletop services Tabletop Scribe, Kazkar, Saga20 | Meeting transcribers Otter, Read.ai | General AI assistants NotebookLM, ChatGPT | Writing it up yourself the benchmark, not a rival |
|---|---|---|---|---|---|
| Spells your characters, places and items the way your table does | ✓Yes | ✓Yes | ✗No | ◑Partly | ✓Yes |
| Keeps every line attached to the character who said it | ✓Yes | ◑Partly | ◑Partly | ✗No | ✓Yes |
| Reads your own lore documents before it writes | ✓Yes | ◑Partly | ✗No | ◑Partly | ✓Yes |
| Writes continuous narrative prose, not bullet points | ✓Yes | ✓Yes | ✗No | ✓Yes | ✓Yes |
| Collects the quotes, the jokes and the grisly bits as their own lists | ✓Yes | ◑Partly | ✗No | ◑Partly | ✓Yes |
| Produces a catch-up recap for whoever missed the session | ✓Yes | ✓Yes | ◑Partly | ✓Yes | ✓Yes |
| The recording never leaves your machine | ✓Yes | ✗No | ✗No | ✗No | ✓Yes |
| You choose the model — or run the whole thing offline | ✓Yes | ✗No | ✗No | ✗No | —Not applicable |
| Change one part later without redoing the rest | ✓Yes | ◑Partly | ✗No | ◑Partly | ✓Yes |
| Runs unattended while you do something else | ✓Yes | ✓Yes | ✓Yes | ✗No | ✗No |
| Yours to read, change and fork | ✓Yes | ✗No | ✗No | ✗No | —Not applicable |
| Works from a browser, with nothing to install | ✗No | ✓Yes | ✓Yes | ✓Yes | —Not applicable |
| What it costs you | Cloud credit by the session, or nothing at all on a local model. No subscription, no account. | A subscription, or credit by the session, depending which. | A monthly fee, whether or not you played that week. | A free tier, then a monthly fee. | An evening. And you would probably do it better. |
Three rows carry the whole argument: where the recording goes, whether attribution is known or guessed, and who decides how it gets written. Everywhere else the tabletop services are genuinely competitive, and the browser row is one they simply win. That is the trade, and it is worth being plain about why it exists — no software can attribute dialogue to the right player unless it is handed something that already knows who was speaking. The only way to know rather than guess is to record each person separately, and that means Discord, a bot and an afternoon of setup you would not otherwise spend.
The last column is the benchmark, not a rival. Someone who actually sits down and writes their session up produces the best version of it there is — the catch is the evening it takes, and that almost nobody does it twice. The middle columns get a fairer hearing than a grid can give them in the full comparison, including the case for using one of them instead.
None of this is about price. It is about the recording staying yours, the names staying right, and you deciding how the story comes out. That is the whole pitch.
How it works
A transcript goes in one end. Each pass hands off to the next, so a failure is recoverable and a long run can be paused and picked up tomorrow.
Your glossary and lore are applied to the raw transcript first, so everything downstream is working from your canon instead of the model's guesses.
The reason this exists
Compares raw against grounded and hands you a list of what it couldn't resolve, rather than guessing quietly.
The narrative pass. Player dialogue stays dialogue; the DM's narration reads as narration.
Quotes, jests and the gorier highlights, extracted as separate lists.
A final review pass, used when you're running the pipeline on a model on your own machine.
Optional. Boils the chronicle down to the short version for posting to the group.
You can send different passes to different models — the cheap ones do the mechanical work, the good one writes the prose. How per-phase routing works →
See it run
Everything above is what comes out. This is the part in the middle: importing a session, pointing it at your campaign's lore, and watching six passes go by.
Being recorded
A video walkthrough is on the way
One real session, start to finish — a recording going in, a chronicle coming out.
Until it lands, the quickstart walks the same path in writing.
Getting a session in
The hard part was never the writing — it's getting four hours of six people talking over each other into text at all. Both roads end at the same grounding pass.
Session audioWhisper, on your PCChronicle
A Discord multitrack, a Zoom export, or a phone on the table. Whisper runs locally, so the audio never goes anywhere. Record each player to their own track and every line in the finished chronicle stays attached to the right character — one track means one person, and nothing has to guess. It also transcribes what was actually said, swearing included.
Wants an NVIDIA GPU, and one module installed.
Session videoYouTube captionsChronicle
Upload the recording unlisted, wait for the automatic captions, bring the transcript back. The transcription happens on Google's hardware instead of yours. Captions carry no speaker labels and censor swearing, so the audit pass asks more questions and the language comes back tamer than your table — but it needs nothing you do not already have.
Wants nothing. No GPU, no Python, nothing to install.
Already have a transcript from somewhere else? Paste it in and skip both. Compare every way in →
What it costs
Tusk's Tomes is free. The only money involved is whatever your chosen AI provider charges for the tokens, and you pay it on the weeks you actually play.
On a Claude or ChatGPT plan you already pay for
$0
On a model running on your own machine
$0
Through one OpenRouter key, cheapest measured routing
~$0.44
Every pass on the strongest Gemini model
~$5
A meeting transcriber, per month, forever, whether you play or not
$15+
Figures are for a three-hour session, priced against the live model catalogue each time this site is published. The full breakdown, per routing →
What leaves your machine
So it's worth being exact about where everything goes. Almost all of it stays on the machine you installed it on.
Your machine
The model you chose
And the honest part: this is one person's project, built and tested on one Windows PC. It is not a company and there is no support desk. The unvarnished version is here.
Questions
[ __ ], already gone by the time you download the file, and nothing downstream can put it back. Whisper transcribes what was actually said. If your table swears, and the way people really talk is part of what you want on the record, the multitrack route is the only one that keeps it.node_modules inside it. Outside that folder it writes one configuration directory in the standard place for your operating system, holding your glossary, routing choices and encrypted keys, and a data directory for sessions if you transcribe audio. That is all, unless you install the transcription module — which adds a Python environment of about 1.5 GB, inside the repository folder. No PATH changes, no registry entries, no system-wide packages, and nothing installed as administrator. The full list, file by file.Start
You need Node 20 or newer and Git. That is the whole requirement. Python and a graphics card matter only if you later want the machine to transcribe audio for you, and you can add them whenever.
Install it I've never used a terminal