> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agentick.xyz/llms.txt
> Use this file to discover all available pages before exploring further.

# Trust at Agentick

> Where your code runs, what leaves your network, and who else sees any of it.

Agentick operates AI agents that do real work in your systems — reading your
tickets, editing your repositories, opening your pull requests, replying in
your Slack. That is only a sensible thing to buy if you can answer three
questions about it, so those three questions are what this section is for.

<CardGroup cols={2}>
  <Card title="Where things run" icon="server" href="/trust/where-things-run">
    Your code, your credentials and your model spend stay on hardware you own.
    What we run, and what we deliberately do not.
  </Card>

  <Card title="What leaves" icon="arrow-up-from-line" href="/trust/what-leaves">
    The complete list of data that crosses from your network to ours, and the
    much longer list of what never does.
  </Card>

  <Card title="Sub-processors" icon="users" href="/trust/sub-processors">
    Every third party involved, what reaches them, and why.
  </Card>

  <Card title="Retention" icon="clock" href="/trust/retention">
    How long anything is kept, including the parts we have not finished
    negotiating.
  </Card>
</CardGroup>

## The short version

**Your source code never reaches our servers.** Agents execute on a machine you
own, using the tool logins already on it. We hold no credential to any of your
repositories — not a token, not a key, not an app installation. We cannot clone
your code, and we could not hand it to anyone who asked.

**Model spend for the actual work is yours, on your key, from your machine.**
Prompts containing your source, and the completions that come back, egress from
your hardware directly to whichever model provider you have configured. They do
not transit us.

**One thing does leave, and it is one sentence long.** The text of a message
addressed to your agent in Slack — or a Linear issue's title and description —
is sent to a classification model so we can decide whether the message is work
at all. Nothing else goes with it: no identifiers, no repository contents, no
diffs, no prompts, no completions. [The detail is
here](/trust/what-leaves), because a reviewer should not have to take a summary
on faith.

## What we would rather you read than discover

This section states the parts that do not flatter us — an isolation boundary
that is weaker than we would like on some hardware, a credential model that
trades blast-radius control for not duplicating your tooling, and a retention
term we have chosen not to pay for yet. They are on [Known
gaps](/trust/known-gaps) and [Retention](/trust/retention).

Finding those during a security review is worse for both of us than reading
them here.

<Note>
  Something missing, or something that reads as evasive? Write to
  [support@agentick.xyz](mailto:support@agentick.xyz). A question a buyer had to
  ask twice is a page we have not written properly.
</Note>