Recording your D&D game with Craig, the Discord bot that gives you one file per player

Have the consent conversation first

Do this before the setup, not after, because it is the part that actually matters and the part most guides skip. Craig's own documentation puts it bluntly: it is both immoral and illegal to record anyone without their permission. Recording law varies by jurisdiction and your table may span several of them.

The full version of that conversation, including why an online game spanning several states is the harder case, is in consent and the law.

The conversation is short and worth having out loud rather than in a text channel. Say what you are recording, why, where the audio goes, and how long you keep it. In practice most tables are fine with "I want to write better recaps and I keep forgetting NPC names." What people object to is finding out afterwards.

Craig is built to make covert recording hard, which is a point in its favour. It renames itself to add "[RECORDING]" to its nickname while it is running, and it will not record at all if it cannot change its own nickname. If someone joins late, the bot in the member list is the disclosure.

• Ask before the first session you record, not mid-session
• Say who can access the audio and for how long
• Agree what happens to out-of-character conversation in the recording
• Give anyone who is uncomfortable a real way to say no

Setting it up

If you have not settled on Discord recording generally yet, the broader Discord recording guide covers the alternatives and when Craig is the wrong choice. This page assumes you have picked it.

Invite Craig to your server from craig.chat. When you invite it, make sure it can send messages, join voice channels, change its own nickname and send you a direct message — the nickname permission is not optional, because that is how it signals recording, and it refuses to run without it.

To start a recording you need the Manage Server permission, or an access role your admin has granted with /server-settings. That catches people out on servers where the DM is not the owner.

1. Join the voice channel your game runs in
2. Type /join — optionally naming a channel, as in /join channel: Recording
3. Check the bot's nickname now reads [RECORDING]
4. Play your session
5. Type /stop when you are done

It is /stop, not /leave

Worth stating plainly because the wrong command is in a lot of older write-ups: the current command set is all slash commands, and there is no /leave. You start with /join and you end with /stop.

Two others earn their place. /note drops a timestamped marker during the session, and Craig writes those notes into the info file that comes with a multi-track download — and into a label track if you take the Audacity project — which is the cheapest possible way to find "the bit where they killed the duke" in a four-hour file later. And /recordings lists your previous recordings, which is how you get a download link back; the next section covers that.

How to download a Craig recording (and the 7-day deadline)

Downloading is where most Craig recordings are actually lost — not to the bot, but to a link nobody can find a week later. On the free tier a recording expires after 7 days, so treat the download as the last step of the session, not a job for later.

The link comes by direct message to whoever started the recording, sent when the recording starts — so it is sitting in that person's DMs before anything can go wrong. If someone else ran /join, they have the link, not you. If your DMs are closed, run /recordings in the server instead. Craig's FAQ is plain that there is no other delivery route; its suggestion is to enable DMs in another server that has Craig in it, or in Craig's own support server.

On the download page you choose a format and wait while Craig processes it. The Multi-track buttons give you a ZIP: Craig's primer says its multiple files are packaged in one ZIP, with one file per speaker inside, each labelled with that person's Discord username. Unzip it on a computer; a phone needs an extractor app. The Single-track buttons give you one file instead, with everyone merged into it. There is no MP3 option — Craig's FAQ answers "Can I export to MP3?" with a flat "No" and says to take FLAC and convert it if you need one.

What to download depends on where the audio goes next. A transcription tool that wants one file cannot use the ZIP, so use a Single-track button — AAC is the smallest — or, if the in-browser mix struggles, open the Audacity project and export one mixed file as MP3 or M4A. If the destination is DM Scribe, a four-hour mix saved as FLAC or WAV can run past its 500 MB upload limit, while an MP3 at around 128 kbps lands nearer 200 to 250 MB. Tier 2 supporters also get Craig's Smart Mix, auto-levelled and mixed on Craig's servers rather than in your browser. And a Tier 3 supporter already has a transcript on the same download page: take the plain-text version, not SRT or VTT — each line already carries the speaker's name, in order — and it can go straight into a transcript import without spending any transcription hours.

1. Find the direct message Craig sent whoever ran /join — or run /recordings if DMs are closed.
2. Open the link, choose a format and wait for it to process: FLAC to keep, AAC if size matters, a Single-track button (or the Audacity project) if you want one mixed file.
3. Download it. A Multi-track download is a ZIP, so unzip it on a computer.
4. For a single file to upload, use a Single-track button (AAC is the smallest), or export a mix from the Audacity project as MP3 or M4A if the in-browser mix struggles.
5. Do it the same night. Free recordings expire after 7 days; supporter tiers keep them for 14 or 30.

The free limits, and what happens when you hit them

On the free tier Craig records up to 6 hours in one recording and keeps the audio for 7 days before it expires. Both numbers matter for a normal game: a four-hour session is comfortably inside the limit, but a long one-shot or a session that runs over is not, and a week is easy to lose to a busy weekend.

When the time limit is reached Craig stops the recording and tells you it has hit the maximum, rather than silently truncating the file. That is the good version of that behaviour, but it still ends your recording mid-session.

The practical advice is to download the audio the same night — the section above walks through it. Supporter tiers, paid through Patreon, extend both limits: Craig's supporter page lists 14-day storage at $1 a month, and 30-day storage with a 24-hour recording limit at $4. Any paid tier can also upload finished recordings automatically to Google Drive, Microsoft OneDrive or Dropbox, which takes the deadline off the table for good.

Multi-track is the default; a one-file mix is free too

This is the part most explanations get backwards, so it is worth being precise. Craig records every speaker to their own file by default — that is the whole point of it, not an option you enable. The Multi-track download is a ZIP containing one audio file per person.

A single-track mix, where everyone is combined into one file, is on the free download page too: Craig's Single-track buttons (FLAC, WAV or AAC) merge every track into one file. The mixing runs in your browser, though, and Craig warns it can be unsupported or take a good amount of memory in some browsers, so try it on a short recording before you rely on it for a four-hour one. The paid version is Tier 2's Smart Mix — "auto-leveled and premixed audio", $4 a month — mixed on Craig's servers instead. So on the free tier you can have one combined file; what you pay for is the levelling, and not doing the work in a browser tab.

Multi-track is genuinely better for a game anyway. You can level the quiet player, cut the person whose dog barked for ten minutes, and drop out-of-character sidebars without touching anyone else's audio.

Which export format to pick

Craig offers more formats than anyone needs, so here is the short version. If you want to edit the session, download the Audacity project — it is FLAC inside an .aup zip and it opens with every track already laid out, which saves you assembling them by hand. If you want files to keep or feed to something else, take FLAC: it is lossless and Craig's own primer recommends it when your software supports it and your connection can handle the size.

Take AAC if the download size matters or you are on a slow connection. It is lossy but it is the most widely supported lossy format, and for spoken word the difference is not what will limit your transcript quality.

Do not reach for WAV out of habit. The WAVs Craig makes on its own servers are reduced quality — ADPCM and 8-bit, under "Show all formats" — and the WAV buttons that build in your browser produce uncompressed files, the biggest on the page. FLAC is the lossless file Craig makes itself.

A correction worth making because it circulates: people often describe Craig's output as per-speaker .ogg files. Ogg formats are available — Ogg FLAC, Opus and Ogg Vorbis are all in the alternate-formats list — but what you get is whichever format you chose on the download page, delivered as a ZIP. Pick FLAC and you get FLAC.

• Editing the session → Audacity project (FLAC in an .aup zip)
• Archiving, or transcribing each speaker separately → FLAC
• One file to upload for transcription → a Single-track AAC download, or a mix exported from the Audacity project as MP3 or M4A
• Slow connection or storage-constrained → AAC
• Skip WAV — the server-made ones are reduced quality and the in-browser ones are the biggest files on the page; take FLAC

Run Giarc as well, not instead

Giarc is a backup clone of Craig, hosted in a different country. The common misunderstanding is that it is an alternative to pick between; it is not. You invite both bots to your server, bring both into the voice channel, and run /join on each. You end up with two independent recordings and two download links.

For a session you cannot repeat — a finale, a character death, a one-shot with a guest — that redundancy is cheap insurance against one bot dropping out. It is still a current, documented part of the product rather than a legacy curiosity.

The other capture path, and self-hosting

Craig also has a browser-based webapp, enabled with /webapp on, where participants join through a link instead of relying on Discord's audio. It captures Opus at 128kbit — better than Discord's own — and Tier 3 supporters ($8 a month) can record 24-bit FLAC through it. Worth knowing if Discord audio quality is your bottleneck.

Craig is open source under the ISC licence and actively developed. Self-hosting is officially documented but Linux-only and genuinely involved — PostgreSQL, Redis, Node, PM2, Docker and a stack of audio tools — so it is a project rather than an afternoon.

Is Craig safe? What it stores, and who can download your audio

Safe is several questions, and Craig's privacy policy answers most of them in plain language — worth reading before you ask five friends to be recorded by it. What follows is what Craig's own pages say, not an audit: nobody here has inspected its servers, and none of this is legal advice.

What it collects: all the audio sent to it, plus the Discord usernames, user IDs and avatars of everyone recorded, and the names of the server and channels — which is how each track arrives labelled with its speaker. Its web services log IP addresses, as most sites do. And it collects nothing at all about someone in the server who never speaks in a recorded channel, never uses a Craig command and never joins through the webapp.

How long it keeps it: audio is normally kept only until the recording expires, though the policy says recordings are sometimes retained longer, for testing or debugging, and that the maintainers ask permission first where they can. Anyone recorded can ask to see their data, or to have it expunged.

Who can download it: anyone holding the link. Craig sends it privately to whoever started the recording, but the policy is explicit that the URL is the key and that keeping it secret is that person's job. Share it with the people who were in the recording and nowhere else.

Who runs it: volunteers. The policy opens with "Craig is not a business. We're not interested in your data." That is reassuring about motive, and also a reminder that no company stands behind its uptime — the real case for downloading the same night, and for running Giarc alongside it when a session cannot be repeated.

And the safety that matters most at a table is social rather than technical. Craig announces itself with [RECORDING] in its nickname and will not record at all if it cannot. Covert recording is what it is built to prevent, which is why the consent conversation at the top of this page comes first.

Now you have six hours of audio and no notes

This is the part nobody warns you about. Recording is solved; the recording is not the thing you wanted. What you wanted was to know what happened, and a ZIP of audio files is arguably worse than notes because it feels like you have the session when you have only deferred the work.

On the free tier Craig gives you audio (a multi-track download adds an info file with any /note markers) and no text of what was said. Tier 3 supporters, at $8 a month, can now download a transcript too — WebVTT, SRT or plain text, made with Faster Whisper — but a transcript is still not notes, and Craig's documentation describes captions and transcripts, not summaries. So for most tables the next decision is still what turns the audio into text, and then the text into something usable.

Running Whisper yourself is free and works well on plain speech. The two things it does badly are the two things a D&D session is made of: it mangles invented proper nouns, because it has never seen your NPC names and will spell them differently every session, and it does not label who is speaking. The fantasy-noun half of that is what campaign-aware transcription exists to fix, and the Whisper guide covers how to work around both yourself — including the fact that OpenAI now offers a separate model that does return speaker labels, if that is the part you care about. The speaker half is also where Craig's multi-track output actually helps — transcribe each speaker's file separately and you have attribution Whisper could not give you, at the cost of doing it several times and interleaving the results by timestamp. A Tier 3 supporter can skip that step: Craig's plain-text transcript is already one file for the whole table, with each line giving the speaker's name and what they said, in order.

Meeting tools are the other obvious option, and their constraint is length rather than quality: Otter caps a single conversation at 30 minutes on its free plan and 90 on Pro, so a four-hour session does not fit one recording until its Business tier. The sourced numbers are in the Otter comparison, and the same session priced across every option is in what it costs to transcribe a session.