Log in Sign up

The human side
of AI at work.

See how AI makes your people feel.
Without anyone at work reading a single chat.

Try it free Start with your team

Free trial takes 3 minutes no credit card

How it works

From a chat to a private dashboard

Capture

The extension or the connector records chats, only while capture is on.

De-identify

Names and numbers are replaced before anything is stored.

Analyse

The model reads each message on our own machine. Nothing leaves it.

Show

Each person reads their own dashboard. HR reads group figures.

The full method
What Custodian reads

166 mental states, one map

Each message is placed on the 3D Mind Model, a published map from social neuroscience. Its three axes are rationality, social impact and valence.

How the model works

Point at the map to read a state.

What people see

What Custodian shows

States of each message

The five most likely of 166, ranked.

Weekly trends

Emotional tone, enthusiasm, exhaustion and stress.

The next turn

Where the conversation is heading, before it gets there.

A safety check on the assistant

Finds 84% of harmful replies in our held-out test.

Personal data removed

Names, emails, numbers, addresses and places, before storage.

Group figures for HR

Averages and shared states, for groups of five or more.

One dashboard for each person, one for HR

A participant dashboard: emotional tone, enthusiasm, exhaustion, stress and safe replies, a trend over three weeks, and the states of the week, the month and the next week.

An example with invented data.

Who sees what

The results belong to the person

Each person sees their own results. Everyone else sees group figures or counts.

The participant

  • Their own states, trends and safety figures
  • Download or erase everything

HR

  • Averages and trends for five people or more
  • Names, chats or one person's results

Custodian

  • How many people take part, under anonymous IDs
  • Chat text or anyone's results
Swiss technology

Built on Apertus, the Swiss open model

Custodian runs on Apertus 1.5, the model of the Swiss AI Initiative from EPFL, ETH Zurich and the Swiss National Supercomputing Centre.

A project from Lausanne, on a Swiss model. The model is open, so anyone can inspect it. It runs on our own machine, and no chat goes to an outside AI service.

ModelApertus 1.5, from the Swiss AI Initiative
ProjectCustodian, developed in Lausanne, Switzerland
LawDesigned with the Swiss FADP, the EU GDPR and the EU AI Act in mind
Results

Tested on 7,364 messages the model never saw

Every score is measured against an automated annotator. It agrees with itself at 0.67, so no model can score higher on these labels.

How to read the numbers
97%Of the highest possible scoreOn the states of each message: 0.65, where the labels allow 0.67.
+46%Over a standard text modelOn the same task, against TF-IDF with ridge regression.
84%Harmful replies foundAbout one in six is missed.
Coding tools

One command for the assistants in your terminal

The connector reads Claude Code, Codex CLI, GitHub Copilot CLI and Antigravity. It sends what you write and the replies, and nothing else from your machine.

The connector
Terminal
For yourself

Fifty messages, free

Talk to your assistants as usual and see what Custodian saw. No card and no email address.

Start the free trial
For a team

Start with one team

From CHF 4 per employee per month. Each person joins by choice and can leave at any time.

See the plans Why it matters