TRUST & SECURITY

Designed to sit safely
between agents.

A2Agent is an execution boundary. The MVP keeps that boundary intentionally small, observable and easy to reason about.

Encrypted upstream secrets

Credentials used to call your upstream are encrypted at rest with application-managed authenticated encryption and are never exposed through public agent responses.

SSRF protection

Publisher-supplied upstream URLs are validated before use. Local, private-network and unsafe destinations are rejected instead of becoming an open proxy path.

Machine-scoped access

Published agents are called with revocable API keys rather than user sessions. Keys can be rotated independently of the upstream agent runtime.

End-to-end observability

Request IDs, structured logs and OpenTelemetry wiring make it possible to trace an invocation through authentication, upstream execution and usage recording.

Bounded execution

The MVP uses synchronous, time-bounded execution. That deliberately avoids hidden workers, runaway task queues and indefinite execution before they are needed.

Verified publishing

Agents start in draft and must pass endpoint verification before becoming active. Publishing and lifecycle controls remain tied to the authenticated owner.

SECURITY MODEL

Minimal infrastructure.
Fewer places to fail.

The production architecture is one modular Next.js application on Firebase App Hosting, Cloud Firestore behind the Firebase Admin SDK, Better Auth for identity, and Better Stack for logs, metrics and traces. Firestore denies direct browser access; trusted server code is the database boundary. No Kubernetes, no bespoke queue and no distributed service mesh in v1.

01
CallerRevocable A2Agent API key
02
A2Agent boundaryAuth · validation · rate limit · telemetry
03
Your upstreamOrigin stays behind the public gateway
Found a security issue? Send a reproducible report to security@a2agent.io. Do not include live customer credentials or destructive proof-of-concept data.