Scylla

Security & Broker

Put a policy between agent intent and execution.

Scylla Broker is the execution gateway for agent actions in Scylla. Your Security Policy determines whether an action is allowed, needs your approval, or is blocked. Keyring provides protected credentials to authorized execution paths without turning those credentials into conversation content.

Broker

The request goes through Broker.

When an agent requests an operation through Scylla, Broker checks the applicable policy before carrying it out. If the operation needs a saved credential, Scylla resolves the protected reference for execution. The agent receives the result it needs to continue, with protected values filtered from supported output paths.

Agent request
Broker policy decision
Allow · Ask · Block
Filtered result

Security Policy

Decide what happens at the operation boundary.

Allow

The operation can proceed when its other access requirements are satisfied.

Ask

Scylla requests approval before the operation runs.

Block

Scylla refuses the operation.

Scope determines available project and reference material. Policy determines what agents may do through Scylla. Windows, repository, and remote service permissions still control what the executing account can access.

Keyring

Store the secret. Give the agent a capability.

Scylla Keyring is an encrypted local store for sensitive connection data. Broker can use those values for approved operations without exposing reusable values to the AI. Unlocking Keyring does not bypass policy.

Terminal & remote access

Bring commands into the policy model.

Broker-controlled terminal and SSH workflows return filtered results for follow-up work. Choose local and remote accounts with permissions appropriate to the task, especially for production systems.

SafeGuard

Help keep secrets out of your repositories.

SafeGuard reduces secret exposure in repository workflows by surfacing findings for review. Keyring protects credential use during execution; SafeGuard addresses the separate risk of sensitive values entering source-controlled work.

Visibility

Review execution and outbound context.

Job monitoring exposes execution output, while payload tracking shows context Scylla sends through tracked AI provider request paths. Remote AI requests still send selected prompts and context to the configured provider.

Scylla’s controls govern its managed execution paths. They complement operating-system and service permissions; they do not turn arbitrary programs into isolated code.