Skip to main content
Connect your agent with the SDK. Its actions stream into the console as metadata only — and the risky ones are governed by policies: rules that allow, flag, or block an action before it runs.

​
How it works

1

Capture

The SDK records each action the agent takes — metadata only, never your data.
2

Detect

Actrail replays each closed session and flags risky patterns.
3

Decide

You confirm what should be governed — from a flag, or from the managed catalog.
4

Enforce

A policy blocks the matching action on the agent’s next run.

​
Where policies come from

Managed catalog

Curated rules Actrail installs when you connect — secret and PII egress, destructive production actions, and per-vendor tool suites for the MCP tools your agents already use (GitHub, Slack, Stripe, Postgres, and more). They arrive in shadow, so you can watch before you block.

Discovered

Actrail flags a risk from your own trails; you resolve the flag into a policy tailored to what your agent actually did.
Either way, a policy starts in shadow — reporting what it would block — and you graduate it to enforce when you trust it.

​
Key terms

Trail

One agent session — the ordered timeline of what it did.

Policy

A rule that allows, flags, or blocks an action.

Managed policy

A curated rule from Actrail’s catalog, installed on connect.

Shadow

A policy that reports what it would block, without blocking.

Enforce

A policy that actually blocks the matching action.

Graduate

Promote a policy from shadow to enforce, once there’s evidence.

​
Get started

Connect Claude Code

Set up in about five minutes.

Using another agent?

Only Claude Code is supported in beta — contact the team for early access.