How Codsworth reads messages
Last updated: October 31, 2026
The short version
- Codsworth only reads messages in channels a server admin turned on. Every channel starts off.
- It reads messages nobody sent to it for four things: learning fixes from support, catching scams, answering crash logs, and crediting helpers.
- Commands, right-click actions and chat don't need that: they only see what you send to Codsworth.
- Servers choose whose messages it learns from, and anyone can turn learning off for themselves on My Data.
- Stored messages are encrypted on disk, kept only as long as listed below, and never used to train AI models.
What reads messages nobody sent to Codsworth
These happen while members talk to each other. The person asking for help doesn't know a bot is there, and a scammer will never mention it, so a command can't replace them.
1. Learning from support
In tickets and help channels with Learn from support on, Codsworth follows how people solve a problem. When the person says it's fixed, the AI reads the conversation with names replaced by "Player" and "Helper", and keeps a short summary of the problem and the fix. The next person with the same problem gets that fix. If a suggested fix doesn't help ("still crashing"), it stops being suggested. Fixes for one server or modpack stay with that server; general game fixes (no names, no server details) can also be suggested on other servers.
- Kept
- Tickets and help threads: until they close or are quiet for 48 hours, at most 14 days. Help channels: up to 6 hours, in memory only. Then the messages are deleted and only the summary stays, without names or user IDs.
- Who decides
- Server admins pick the channels, and whether to learn from everyone, everyone except people with a role, or only people who opt in with a role. Every member can turn it off for all servers on My Data.
2. Catching scams
Links are checked against known scam sites. Shortened links are followed to where they really go, and links that still look suspicious are looked up with IPQualityScore (only the link is sent, not the message or who posted it). Images, in channels with image scans on, are read by the AI for scams like fake crypto giveaways and fake support pages, which a word filter can't see. When something is found, the server's moderators get an alert.
- Kept
- Only messages that look like scams: a report for the moderators with the message text, links and what the image shows. A scam image is posted only to the moderators' alert channel on Discord, not stored by Codsworth. Everything else is checked and forgotten.
- Who decides
- Server admins turn link checks on for the whole server, and image scans per channel. They apply to everyone, like any other moderation.
3. Crash logs and load orders
When someone posts a crash log, a pastebin link to one, or a load order, Codsworth replies with the likely cause and fix.
- Kept
- The crash log, up to 180 days or until you delete it. Crash statistics (where the game crashed and which mods were near it) are kept to improve analysis; they contain no user data.
- Who decides
- Server admins choose the channels. It's on in a channel once they add it, and they can turn it off there.
4. Crediting helpers
When someone thanks another member ("thanks @Ann, that fixed it"), the helper gets credit on the server's leaderboard. If nobody is named, Codsworth looks at the last few messages to work out who helped.
- Kept
- Only counts: how often someone was thanked and thanked others. Not the messages.
- Who decides
- Server admins choose the channels. It's on in a channel once they add it, and they can turn it off there.
What only sees messages sent to Codsworth
- Slash commands like
/faq,/healthand/summarize. - Right-click actions: Analyze crash log and Ask Codsworth on a message you pick.
- Chat: messages that mention Codsworth, reply to it, or are sent to it directly. It doesn't join conversations it wasn't asked into. Conversations are kept 30 days for context. In a channel where the server turned on shared channel memory, Codsworth also sees what others there said to it in the last 12 hours, and the names of who posted there recently (kept in memory only), when it answers.
- Voice chat: only after someone uses
/voice join, and only in that voice channel: what people say there goes to OpenAI so Codsworth can answer out loud. It isn't recorded or saved. - Corrections: when you reply to Codsworth that something it said was wrong, it can learn from that, unless you turned learning off.
Your choices
As a member
- Type
/privacyin any channel to see what Codsworth reads there, and turn learning from your messages off with one button. - Or turn it off on My Data. Either way it applies on every server, and what Codsworth was still following from you is dropped right away.
- On servers that use an opt-in or opt-out role, add or remove that role yourself.
- Download or delete your data at any time on My Data.
As a server admin
- Codsworth does nothing in a channel until you add it on the Channels page, and you choose what it does there. Scam link checks are one switch for the whole server, off until you turn them on.
- Choose whose messages are learned from on the Config page.
- Remove Codsworth and its server data from the dashboard.
How it's protected
- Stored messages, crash logs and learned fixes are encrypted at rest.
- Message text goes to our AI providers (OpenAI, or Anthropic if a server picks a Claude model) only to answer, summarize or check for scams. They don't use it to train their models.
- We don't train or fine-tune any AI model on your messages, and we never sell them.
The Privacy Policy has the full details, including the legal basis and your rights. Questions: message zerratar on Discord, or see Support.