How it compares

I tried most of the meeting tools before writing my own, so that part is a fair account rather than a sales pitch. They're genuinely good software, just built for meetings, and a session isn't a meeting.

The tabletop-native tools are a different matter and deserve saying plainly: there is a real category of them now, several arrived after this project started, and I have not sat down with all of them. What's below is what they are built to do, not a ranking of how well they do it.

Being fair about it

The tabletop-native services are the comparison that actually matters, and being vague about them would be a way of avoiding it. They're built by people who play. Several advertise speaker attribution and mean it. Most will take a recording and give you a recap without asking you to install anything, which is a genuine advantage over this project and not a small one.

Two things separate them from what's here, and neither is about quality.

The first is where your session goes. They are services, so the audio of your friends talking for four hours is uploaded and processed on hardware you don't control. That may be completely fine with you — but it's a decision worth making deliberately rather than by default, and it's the one thing none of them can offer differently, because being a service is what they are.

The second is how a line ends up attached to a player. No software can attribute dialogue accurately unless it's fed something that already knows who was speaking. Hand over a single mixed recording — an mp3 off a phone in the middle of the table, or a bounce out of OBS — and every name on every line is an inference. Modern diarisation is good, and on a clean recording with distinct voices it's often right. It is still a guess, it degrades exactly when your table is at its best (everyone talking at once), and it cannot know that the voice it's tracking is the one playing the paladin.

Tomes sidesteps that rather than solving it: Craig records each player to their own file, each file is transcribed separately, and attribution arrives with the recording instead of being reconstructed afterwards. The cost is that you have to be playing on Discord with Craig in the server. On its own YouTube path, Tomes has exactly the same limitation as everything else in this table — one mixed recording in, guesswork out. That path exists because not everyone has a GPU, not because it solves anything.

Otter and Read.ai do what they set out to do perfectly well. Point them at a standup and you'll get a tidy transcript and sensible action items. Point them at a session and you get a transcript with the names spelled wrong and a summary that says the group "discussed next steps". Not their fault — nobody built them for this.

NotebookLM is the closest thing to real competition, and I'd say so even though it's not what I use. Feed it a clean transcript and the output is strong. What it won't do is transcribe your Discord recording per-speaker, correct names against a glossary you maintain, keep character attribution through to the finished text, or run without uploading anything. If you're happy doing the front half by hand and you don't mind the upload, it's a perfectly reasonable option.

A human transcriber produces the best result available, by a distance. It also costs more per session than a year of API credit. If you can afford it, genuinely, do that instead.

Craig isn't a competitor at all — it's the other half. It records beautifully and does nothing else, which is exactly right. This project is the part that was missing after it.

So what's actually different here

Narrowly, three things. Note that "it knows your characters' names" is no longer one of them — the tabletop-native services ground against your lore too, and claiming that as a differentiator would be dishonest.

1. It runs on your machine. The audio never leaves it, and with a local model nothing leaves it at all. This is the only row in the table where nothing else comes close, and it's the reason to pick this over a service that is easier to start using.
2. Attribution is known, not inferred. Per-speaker Craig tracks are transcribed separately, so who said what is an input rather than a conclusion. Everything downstream — quotes, jests, the chronicle — inherits that instead of re-deriving it.
3. You decide how it comes out. Which model writes each phase, what the narrator sounds like, what the prompts say, how long the recap runs. It's MIT, so if the answer you want isn't in the settings it's in the source.

If none of those matter to you, one of the tools above will serve you better and cost you less effort. That's a genuine recommendation, not false modesty. The version of this project that tells you it wins on every axis is one you should trust less, not more.