Product
What Custodian does
Four parts, all in use today.
The dashboards
What each person and HR see
An example with invented data.
Capture from the assistants your team uses
A browser extension for ChatGPT, Claude, Gemini and Microsoft Copilot. A connector for Claude Code, Codex CLI, GitHub Copilot CLI and Antigravity.
- Off until the employee switches it on
- Only the conversation: no other sites, no code, no files
- The employee can read, remove or exclude a chat before it is sent
A private dashboard for each person
Each participant sees their own states and weekly trends in a private dashboard.
- States of the week, the month and the next message
- How often the assistant's replies were safe
- Download or erase everything, without asking anyone
A safety check on the assistant
Every reply is rated safe, borderline or harmful, including the dismissive or hostile replies that general filters miss.
- Finds 84% of harmful replies in our held-out test
- Misses about one harmful reply in six; no human reviews the ratings
A group view for HR
How the people who take part are doing as a group, to decide on tools, workload and training.
- Groups of five or more only, for the organisation and for each team
- A state or a week shows only when five or more people share it
- No names, accounts, chats or individual scores
Limits
What Custodian does not do
The limits of the product today.
- No diagnosis. Custodian is not a medical device.
- No real-time alerts. Dashboards refresh every month, week or day, by plan.
- No link to HR systems, and no individual results for anyone but the participant.
- No EU AI Act conformity assessment yet. We go through the classification with you first.
See it on your own conversations
Try it yourself in a few minutes, or start with one team.