Back to Gap Map Product / UX / Agent identity & authorization

Caracal

Garudex Labs
open core / Overall score: 2.6

Caracal is a self-hosted authorization service that gives AI agents short-lived, signed mandates instead of credentials. A policy service decides each request before it runs, a gateway injects the real upstream credential at call time, delegated mandates can only narrow, and every decision lands in a tamper-evident audit chain. Garudex Labs builds it, with TypeScript, Python and Go SDKs.

Openness

4 high confidence
4.0
license
Apache-2.0 (Community Edition: policy service, gateway, coordinator, audit service, console and SDKs)
source
public
core features withheld
yes — the NOTICE says the repository holds the Community Edition only and that Enterprise features are developed, licensed and distributed separately; the Enterprise Edition adds managed multi-tenancy, a hosted management UI and SAML and OIDC single sign-on with SCIM provisioning
commercial
Caracal Enterprise Edition (a managed deployment sold under a written commercial agreement)

Caracal's published code is a complete Community Edition, with the policy service, gateway, audit service and SDKs all running on the user's own infrastructure. Garudex Labs keeps its Enterprise Edition outside the repository under separate terms, and that edition adds managed multi-tenancy, a hosted console and single sign-on with SCIM provisioning.

Adoption

1 medium confidence
1.0

Adoption is read from monthly downloads of Caracal's Python and TypeScript SDKs, caracalai-sdk on PyPI and @caracalai/sdk on npm, which agents use to obtain and present mandates. The services themselves install from Docker images through an install script, which no registry counts, so the figure shows SDK use rather than running deployments.

Capability

5 medium confidence
5.0

Caracal implements every part of agent authorization itself: each agent run gets its own identity, every request is decided against policy before it runs, delegation can only narrow, risky actions wait for an approver, and decisions are kept in a tamper-evident chain.

Verified 2026-11-20, except some scores, still awaiting re-confirmation