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

# Security model

> Understand workspace isolation, credential boundaries, and display-safe public results.

The public API is the canonical boundary for SDK, CLI, and MCP integrations. These clients do not call the database, cloud browser daemon, or provider control plane directly.

## Workspace isolation

The workspace is derived from the authenticated API key. Requests cannot select another workspace, and resources are returned only when they belong to the key's workspace.

## Display-safe resources

Public run resources include status, timestamps, safe outputs, summaries, and bounded error fields. They intentionally exclude:

* draft Blueprint graphs and authoring state
* raw browser observations or screenshots
* internal node events and tool arguments
* cookies, website credentials, and secret values
* provider object identifiers and control-plane details

Blueprint event polling returns only a bounded `run.snapshot` representation.

## Credential handling

* Create keys in authenticated Developer settings.
* Grant least-privilege scopes.
* Store secrets in a server-side secret manager.
* Rotate with a bounded overlap window.
* Revoke immediately when a consumer is retired or compromised.
* Never send a key through task text, workflow variables, metadata, URLs, or logs.

## Cloud execution

Cloud browser work runs in an isolated environment that cannot see local cookies, files, or logged-in browser sessions. If a task requires a user's existing local session or direct human control, use the Infragrid desktop product instead of the public cloud API.

## Reporting an issue

Share run and request identifiers, not secrets. See [Errors and retries](/production/errors-retries) for the safe diagnostic checklist.


## Related topics

- [Infragrid overview](/overview.md)
- [MCP server](/tools/mcp.md)