Authority for AI agentsAuthority for AI agents

How it works

Five steps from request to evidence

AI agents request scoped authority for protected actions. Follow policy evaluation, authority issuance, action enforcement, and tamper-evident audit.

Request actionEvaluate policyIssue authorityEnforce actionRecord evidence

Step 1 of 5

An AI agent requests only the Resource scopes required for a protected action.

Identities and Applications

Step 2 of 5

STS evaluates active Policy, Session, Delegation, Resource, scope, and Approval context before approving authority.

Policies and Policy Sets

Step 3 of 5

STS signs a short-lived Mandate bound to the accepted Application, authority anchors, Resource, and scopes.

Mandates

Step 4 of 5

Gateway or an in-process verifier checks the Mandate and revocation state before the protected action runs.

The Gateway

Step 5 of 5

Audit appends hash-chained evidence that correlates the policy decision and protected action result by request and trace identifiers.

Audit and Request Traces

Try it nowcurl -fsSL https://raw.githubusercontent.com/Garudex-Labs/caracal/main/install.sh | shthen caracal up

Policy-approved actionsActive Policy evaluates the agent, Session, Resource, scopes, Delegation, and Approval context before authority is issued.Delegation can only narrowA child can never exceed its parent’s scopes, Resources, lifetime, or hop budget.Instant revocation at GatewayGateway-routed actions perform fresh authority checks and reject revoked Sessions immediately; direct verifiers remain bounded by propagation and Mandate TTL.Tamper-evident auditHash-chained, append-only records correlate every authority decision with the protected action result.

Find your path

Where are you starting from?

I'm new hereI'm buildingI run the platformI want to contribute

Install the released runtime, protect one call, and verify the matching decision and action result before reading deeper reference material.

What is Caracal?

• Install Caracal
• First protected call
• Caracal mental model
• Add SDK to your app
• First-run troubleshooting

Map your Application and Resources, choose an enforcement boundary, integrate an SDK, and validate allow, deny, and revoke behavior.

Model your application

• Author policy data
• Protect your first real API
• Multi-agent delegation
• Tail the audit stream
• TypeScript SDK

Choose a supported deployment shape, configure secrets and network boundaries, and rehearse monitoring, recovery, and upgrades.

Operations guide

• Deploy with Docker Compose
• Deploy with Helm
• Harden production
• Monitor health and metrics
• Troubleshoot by symptom

Set up the repository, learn the component boundaries, run focused validation, and follow the documented review process.

Contribute to Caracal

• Set up locally
• Understand the architecture
• Review the threat model
• Make a change
• FAQ

Caracal

Authority, not credentials, for AI agents.

Star on GitHub170

Adoption

Get startedReview OSS readinessOperate Caracal

Community

BlogsVlogsGitHub repositoryWeekly Dev Call · Thu 9pm IST

© 2026 Garudex Labs. All rights reserved.

Caracal is licensed under Apache-2.0.