
🐾 Authority, not credentials, for AI agents: policy-approved actions, delegation that can only narrow, instant revocation, tamper-evident audit.
Authority, not credentials: Policy-checked, attributable, revocable, provable AI agent authorization.
"Gateway-mediated agents never hold upstream credentials. Every action is policy-approved before it runs, scoped to exactly what was delegated, revocable in one call, and recorded as tamper-evident evidence."
AI agents are entering production with long-lived API keys in their environment, broader access than any task needs, and no answer to "which agent did this, under whose authority?"
One prompt injection, leaked key, or runaway loop turns an assistant into an incident. Security reviews block launches. Auditors have nothing to inspect.
Existing tools weren't built for this: identity providers register agents but never see their actions, secrets managers hand the credential to the workload, and API gateways route traffic without deciding anything. Caracal is the missing control plane. It decides what an agent may do before every action, proves what it actually did, and shuts it down instantly when something goes wrong.
For gateway-mediated calls, agents never receive upstream credentials. They carry mandates: short-lived, signed grants of authority that can only shrink as work is delegated. Caracal's gateway injects the real credential at call time, so the upstream credential never enters the agent process.
flowchart LR
A[Agent] -- requests authority --> STS[Policy decision<br/>default deny]
STS -- short-lived mandate --> A
A -- presents mandate --> GW[Caracal Gateway]
GW -- real credential injected --> API[Tool / API / MCP server]
STS -. every decision .-> AUDIT[(Tamper-evident audit)]
GW -. every call .-> AUDIT
| Capability | Outcome for your team |
|---|---|
| No standing secrets | For gateway-mediated calls, upstream credentials stay at the edge and are injected only for the request. |
| Delegation that only narrows | Agents hand work to sub-agents, never with more access than they hold. Least privilege is enforced, not requested. |
| Decisions before actions | Default-deny policy evaluates every request before it reaches a resource, not flagged in a log afterwards. |
| One-call kill switch | Revoke an agent and everything it started loses access at once. No waiting for tokens to expire. |
| Human approval for risky actions | Approval gates hold high-risk operations for an authenticated approver before they execute. |
| Evidence, not just logs | An append-only, tamper-evident trail of every decision, exportable for SOC 2, EU AI Act, NIST AI RMF, and OWASP Agentic reviews. |
| Alternative | What's missing |
|---|---|
| Identity providers | Register agents; never see their actions. No per-call decisions, no delegation limits, no evidence. |
| Secrets managers | Hand the credential to the workload. A compromised agent is a leaked secret. |
| API gateways | Route and rate-limit; carry no authority model and no delegation semantics. |
| Building it yourself | A proxy + vault + policy engine gets you plumbing, not narrowing delegation, revocation propagation, or a tamper-evident audit chain. Then you maintain it forever. |
Caracal is standards-native: OAuth 2.0 token exchange (RFC 8693), OPA policy, and MCP integrations. It fits the stack you already run instead of replacing it.
Per-request authorization usually means a latency tax. Caracal's is smaller than most gateways spend on plain proxying — measured end to end on a single 4-vCPU VM running the packaged stack with stock configuration: