
Argent decides with AuthZed SpiceDB
If you model permissions as relationships in SpiceDB, Argent can ask your cluster whether an AI agent holds a permission on what it is about to call. Only a definite yes lets the request through.
What Argent does with SpiceDB
Keep the policy engine your team already runs. Argent asks it before any agent request goes out.
NIST's zero trust architecture, SP 800-207, puts it plainly: "Authentication and authorization (both subject and device) are discrete functions performed before a session to an enterprise resource is established." (NIST SP 800-207). Argent applies that to AI agents on every request: a policy engine decides, not the agent.
Most tenants use Argent's built-in OPA engine. A tenant that already writes its authorization rules in its own engine can route Argent's decision there instead, so one set of rules covers people, apps and AI agents.
SpiceDB describes itself as "an open-source, Google Zanzibar-inspired database system for real-time, security-critical application permissions" (AuthZed).
- One permission check per request. Argent asks: does this agent, as a subject, hold this permission on this resource?
- Agent and path. The subject id is the agent Argent identified; the resource id is the request path.
- Only a definite yes. The request passes only on
PERMISSIONSHIP_HAS_PERMISSION. Anything else is a deny. - Validated end to end. This connector is live-validated: a real allow, a real deny and an unreachable cluster that fails closed.
How it connects
Argent calls SpiceDB's HTTP API. Your schema and relationships stay where they are.
| Direction | Argent calls your engine once per request, at the endpoint you give |
| Call | POST <endpoint>/v1/permissions/check on the SpiceDB HTTP gateway |
| Sign-in | Always the SpiceDB preshared key; Argent refuses a connector without one |
| Certificates | Certificate checks stay on. A self-signed engine is trusted by pinning its CA certificate, never by turning verification off |
| Network | Argent refuses private, loopback, link-local and cloud metadata addresses unless an operator has allowed that network range; the check runs when you save, with a reason |
| Credential | Stored as a reference in Argent's secret store, never in the connector settings, and dropped when the decision is done |
| Consistency | Fast by default (minimize latency), or fully consistent when a stale grant is not acceptable |
What your schema sees
Three names you choose, two values Argent fills in.
| Setting | Example | Filled in at request time |
|---|---|---|
| Subject type | user, nhi or serviceaccount (default user) | the agent's identity |
| Resource type | endpoint, model or route | the request path |
| Permission | call, invoke or view | checked as named |
No prompt, request body, key or credential is ever sent to SpiceDB. Because the check runs before the request body is read, Argent's own model check still runs on the body afterwards, so a model nobody approved is refused as well.
What you configure
One connector, in the console. Your relationships decide.
- Model it. Add the subject and resource types and the permission to your SpiceDB schema, and write relationships that grant agents what they may call.
- Open Policy Engines. In Integrations, open Policy Engines and choose AuthZed SpiceDB.
- Enter the cluster. The HTTP endpoint, such as
https://spicedb.internal:8443, plus the resource type, permission and subject type. - Give the key. The reference to the stored preshared key.
- Choose consistency. Leave it fast, or set fully consistent for decisions that must see the latest grant. Optionally pin your CA certificate and set a time limit up to 30 seconds.
Test the connection from the same page. Once the connector is active, every request from that tenant's agents is decided by your engine. Connecting an engine is an administrator's job: the Integrations page is gated by console role, like every console page.
Fail closed, every time
A policy engine Argent meant to ask but could not reach never allows anything.
- Timeout is a deny. Each decision has a time limit, 3 seconds unless you set another, and never more than 30 seconds. No answer in time means the request is refused.
- Errors are denies. A network error, a status other than success, or a reply Argent cannot read is a deny.
- No quiet fallback. Argent never falls back to its built-in engine when yours fails. Your engine may be stricter than the default, so substituting the default would be a way around your policy.
- Named reasons. A refusal from your engine is recorded as
pdp.external.deny, an engine that failed to answer aspdp.external.error, and one Argent could not build or look up aspdp.external.unavailable. The caller gets a 403; your engine's reply body and headers are never passed back to it. - Signed record. Every refusal goes on Argent's signed record with its reason, so an auditor can see which engine said no and why. See One-Click Audit Evidence.
- One tenant, one engine. The engine is chosen per tenant from the authenticated request, never from anything the caller sends. A tenant with no engine connected uses Argent's built-in OPA engine.
Where this fits
Pages that go deeper.
- Relationships for documents too: retrieval-time authorization for RAG runs on OpenFGA.
- Capabilities: Model Control, Agent Permissions, Watch Before You Enforce and all capabilities.
- Goals: model control and routing, agent identity and ownership, audit and accountability and all goals.
- Use cases: controlling agent tool access, securing AI coding agents and AI in regulated and air-gapped environments.
- Where it runs: air-gapped on-premises, private cloud or your public cloud VPC. Compare them on deployment options.
- Industries: Financial Services, Insurance, Healthcare, Public Sector and all industries.
- More integrations: every system Argent connects to, grouped by type, on the integrations page.
Want this running on your own use cases?
All downloads