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

# Network controls

> What the agent's computer can reach, and how credentials cross the wire

Every request leaving an agent's computer goes through Skydive's transparent proxy. Nothing reaches the internet directly. That single chokepoint is what makes the credential model work and what gives your workspace a place to govern outbound access.

## The proxy holds the credentials

When an agent calls a connected service, it sends the request without a token. The proxy matches the request against the rules for that agent and attaches the credential on the wire.

The agent never sees the secret. It cannot print it, log it, paste it into a conversation, or leak it through a prompt, because it was never in the sandbox to begin with.

<Card title="How this works in detail" icon="shield-halved" href="/docs/integrations/permissions">
  Scoping, revocation, and why the design is shaped this way.
</Card>

## Rules are matched per agent

Proxy rules match on hostname, path, and method, and they are attached per agent. An agent gets a credential only for the rules it has been granted, so connecting GitHub for one agent does not hand GitHub to every agent in the workspace.

A workspace can override a rule for its own tenancy or substitute its own values, without changing anything for anyone else.

## Governing what leaves

Because every request crosses the same boundary, outbound access can be governed centrally rather than agent by agent. That is how workspace-level controls over which services agents may reach are enforced, and how usage is attributed back to the agent and workspace that caused it.

<Note>
  Control-plane traffic is never blocked: an agent that cannot reach a model can still report the problem, checkpoint its work, and tell you why it stopped. A cut-off agent that also goes silent is worse than a cut-off agent.
</Note>

## What this means in practice

* **Secrets stay out of the sandbox.** Store them once, and the agent uses them without holding them.
* **Access is revocable in one place.** Remove a connection and the next request has no credential attached, with nothing to clean up inside the agent.
* **Reach is auditable.** Outbound calls are attributable to an agent and a workspace.

<CardGroup cols={2}>
  <Card title="Secrets" icon="lock" href="/docs/integrations/secrets">
    Storing credentials the agent can use but never read.
  </Card>

  <Card title="Enterprise controls" icon="building" href="/docs/workspaces/enterprise">
    SSO, audit logs, and workspace-level policy.
  </Card>
</CardGroup>