Insights

Know where the time goes.

The bill tells you what the week cost. This tells you where it went: which repository, which people were in it, and how much of the work the model had already seen before.

Book a demo

What Selan is, in one go. Your company already pays for several AI accounts and models. Selan puts them in one setup, shows what each person spends, lets you set limits, and removes secrets it recognises before a request is sent.

This page is about the time.

The same week, read two ways.

Hours by repository, each bar split by the people who were in it. Then the same hours by person. One answers where the work went, the other answers who was doing it.

A duration here is how long a model was working for somebody, not how long they sat at a keyboard.
What it records is the repository, who ran the session and what it cost.
Never prompts, never transcripts, never tool names, never the files anybody touched.

Four repository rows, selan-proxy, selan-website, selan-control and selan-cli, each bar split into per-person segments, with token totals from 111.4M down to 36.5M

How much of the context was paid for twice.

Every request carries the conversation so far. When the model has seen that text before it comes back out of the cache, and when it has not, you pay to process it again. This says which repositories are on the wrong side of that.

Healthy at 75% or better, Watch between 50 and 74, Poor below that, written beside the bar as well as coloured.
Writing something into the cache costs about ten times what reading it back costs, so a repository that keeps missing is a repository quietly spending money.
It is a property of that repository's context, not a score for the people in the section above.

The scale: Poor 0 to 49 percent, Watch 50 to 74 percent, Healthy 75 to 100 percent

The four repositories against that scale: three Healthy at 85, 89 and 90 percent, and selan-cli flagged Watch at 63 percent

Off until an owner turns it on.

This is the part of Selan that looks closest to watching people, so it is the part that ships most firmly switched off. Nothing is collected until somebody at your company decides, and the decision is recorded with their name on it.

The owner's view. Engineers can run the accounts without reading anybody else's week.
Recording which repository a session belonged to is its own separate switch.
Every switch carries the day it was moved, and who moved it.
There is no backfill. The first day it collects is the day it was turned on, and not one day earlier.

Three groups of switches, each stamped with its last change: Secret scanning on at 23 Aug 15:12, Insights on at 23 Aug 09:31, Fail over during an Anthropic outage on, and Claude Code attribution off

How it works

An owner turns it on. Until then this page is empty and nothing is being collected.
Your team works as usual. Nothing to install, nothing to tag, nothing to fill in.
You read the week by repository, by person, and by how much context came back out of the cache.

What to know

This is not a productivity score, and we would rather you did not use it as one. It measures how long a model was working and how warm a repository's context was, which is a fact about the work, not about the person.
The repository name is the name only. Never a path, never the contents.
There is no backfill, so turning it on tells you about the weeks that follow, not the weeks that went before.
The exact list of what is stored, and the one case where a provider's reply is written down, is on Trust.

The other three

Govern

See one week of your own.

Half an hour with the people who built it. Turn it on, work for a week, and read what comes back.