Alter

Alter's pitch is narrower than "auth for AI" and more useful for it. It is a proxy, not a user directory: every request an agent makes to a provider passes through vault, policy, identity, and audit before it leaves. That single choke point is what makes the audit trail credible — the log entry is generated by the same hop that injected the credential, so you get action, agent, end user, policy decision, and approver in one place rather than stitched together from three systems. The credential handling is the part most teams will feel first. OAuth tokens and managed API keys are stored once, injected per request, and never handed to your process — which means a prompt-injected agent has nothing to exfiltrate from its own context, and `.env` sprawl across fourteen agents collapses into one vault. Policy is default-deny with parameter-level controls, so `refunds:write` on a billing agent can be parked for a named human while `calendar.readonly` flows straight through. IdP webhook offboarding cascading across agent keys and grants is the detail that sells this to a security lead; it means one HR action actually revokes access everywhere. The strictness cuts both ways. Fail-closed authorization means an unevaluable policy denies the request — correct for production, occasionally annoying while you are still writing rules. Audit retention is 30 days on Free and unlimited on paid plans, so any compliance-flavored use pushes you off Free quickly. SOC 2 is in progress, not complete, so buyers who need a finished report in hand today should plan around that; the trust center publishes the live state of the program, which is the right posture but not the same as an attestation. Economically, metered-by-call suits agent workloads that fan out — you pay for 100K proxied provider requests on Starter at $50/mo, and SDK management calls (sessions, grant listing, identity resolution) don't count. But high-fan-out pipelines should model the $0.50 per 1K overage before committing; a few thousand proxied calls per session adds up. The Free tier is genuinely complete on features (OAuth, full policy engine, OIDC identity providers, unlimited users and agents) but pauses at 10K calls instead of billing over, which makes it a real evaluation tier and not a production one. Where it doesn't fit: solo builders hitting one API with their own key, teams whose agents only touch first-party infrastructure, and anyone shopping for employee SSO — this is agent authorization, not human login. Against a hand-rolled env-var-plus-shared-key setup, the argument is straightforward: you are going to build the vault, the policy check, and the audit log anyway, and getting fail-closed semantics and SIEM streaming right is weeks of work.