API Key Management

Illustrative mutation fragments

The management mutations on this page are non-standalone fragments. Before adapting them, follow this site's Example Quality Contract and complete Quickstart lifecycle. Do not treat HTTP success or management readback alone as runtime proof.

Use the AI Gateway key API to create, inspect, rotate, and revoke workload credentials. On a mode: 4 rule with api_key_auth: required, the request path enforces key validity, model allow-lists, per-key request rate, and the configured quota ladder. This declaration is independent of sse_mode and pd_disagg_mode.

Request path

flowchart LR
    C([Client]) --> K{"X-Api-Key<br/>valid?"}
    K -->|no| R401([401 invalid_api_key])
    K -->|yes| M{"Requested model<br/>allowed?"}
    M -->|no| R403([403 model_not_allowed])
    M -->|yes| L{"Key and tenant<br/>rate buckets allow?"}
    L -->|no| R429([429 rate limit])
    L -->|yes| B["Inference backend"]

    style B fill:#e8f5e9,stroke:#43a047
    style R401 fill:#ffcdd2,stroke:#e53935
    style R403 fill:#ffcdd2,stroke:#e53935
    style R429 fill:#fff9c4,stroke:#f9a825

API-key configuration uses an independent PostgreSQL data-plane store. It does not depend on the management user service. Management operations use a control-plane bearer identity when management authentication is configured; inference requests use X-Api-Key and never use the management bearer token.

Development-source behavior

The independent key store and authentication-plane separation described on this page are implemented in the current development source but have not completed release qualification. Confirm the flags and API schema exposed by the exact image you deploy.

The service declaration, not SSE or P/D, activates authentication

Omitted api_key_auth preserves a backend-owned X-Api-Key. Explicit disabled strips that header without validating it. required validates and strips it. If a required rule has no usable key store, the request fails closed with 503 policy_store_unavailable; an unknown key returns 401 invalid_api_key. Prove those two failures separately and require a zero backend receipt delta before exposure.

A key store does not protect the management listener

The key lifecycle routes are registered independently of --userservice. If none of the management authentication modes (--userservice, --oauth2, or --manualtoken) is enabled, the current authorizer grants management requests unrestricted access. Before exposing port 11111, enable one management authentication mode and verify an unauthenticated key-creation request returns 401.

Security rules

• Use TLS for both control-plane and inference traffic outside an isolated lab.
• Create one key per workload and environment, with the smallest model list and rate limits it needs.
• The create response is the only place that returns the raw key. List and get return metadata only.
• Never put a raw key in a URL, log, metric label, Git repository, screenshot, or support ticket.
• Prefer protected header files or a secret-aware client over credentials in command arguments.

Prepare a management header file:

export CONTROL_API="https://gateway.example.com/netlox/v1"
install -m 600 /dev/null ./control-plane.headers
printf 'Authorization: Bearer %s\n' "$CONTROL_PLANE_TOKEN" > ./control-plane.headers

Endpoints

Method	Path	Purpose
POST	/config/ai/apikey	Create a key; raw value is returned once
GET	/config/ai/apikey?tenant_id=...	List key summaries, optionally by tenant
GET	/config/ai/apikey/{key_id}	Read one key summary
PATCH	/config/ai/apikey/{key_id}	Update model access, enabled state, RPS/burst/TPM, and priority/objective fields; raw-middleware route
DELETE	/config/ai/apikey/{key_id}	Permanently delete a key
POST	/config/ai/tenant/ratelimit	Set tenant RPS and token quotas
GET	/config/ai/tenant/ratelimit/{tenant_id}	Read tenant limits

Create a key

Only tenant_id is required, but a production key should be explicitly scoped:

curl --fail-with-body --silent --show-error \
  --request POST "$CONTROL_API/config/ai/apikey" \
  --header @control-plane.headers \
  --header 'Content-Type: application/json' \
  --data '{
    "tenant_id": "team-a",
    "name": "chat-service",
    "allowed_models": ["example-chat-model"],
    "rate_limit_rps": 5,
    "burst_size": 10,
    "tokens_per_min": 0,
    "enabled": true
  }' > ./new-key.json

jq '{key_id, raw_key_present: (.raw_key | type == "string")}' ./new-key.json

Expected result: 201 Created, with key_id and raw_key. Move the raw value directly into your secret manager, then securely remove the temporary file.

Field	Type	Meaning
tenant_id	String	Owning tenant; required
name	String	Non-secret operator label
allowed_models	String array	Exact model identifiers this key may use
rate_limit_rps	Integer	Per-key requests per second; 0 disables this limit
burst_size	Integer	Per-key request bucket capacity
tokens_per_min	Integer	Positive values participate in pre-admission reservation and usage settlement on supported token-accounting paths; nonpositive values leave this dimension unlimited. Insufficient reservation returns 429 before backend delivery
expires_at	RFC 3339 timestamp	Optional key expiry
enabled	Boolean	Defaults to enabled when omitted
priority_class	String	Source contract class: standard, critical or sheddable
allowed_objectives	String array	Source contract objective allow-list
default_objective	String	Source contract default objective name

The companion PATCH schema accepts priority_class, allowed_objectives and default_objective as well as the model, enabled and quota fields. These source contracts are described in network identity and objective API contracts; they are not a statement of GAIE/llm-d deployment qualification.

An empty model list records no model restriction. Prefer an explicit list when the workload should use only known models.

The primary Swagger documents per-key admission reservation and usage settlement. Admission estimates prompt tokens and uses declared max_tokens, or a configured completion default when omitted; settlement reconciles the reservation with usage. A request can therefore be refused before its first backend delivery with 429 token_quota_would_exceed and Retry-After. Post-response debt can also affect later requests. These are related stages, not a post-response-only guarantee. Cancellation, missing usage and the minute window require their own checks. Source support and CPU functional tests do not qualify every protocol, engine, HA pair or released product.

Use Gateway-generated keys and retain raw_key only from the create response.

List and inspect without exposing the secret

curl --fail-with-body --silent --show-error \
  --get "$CONTROL_API/config/ai/apikey" \
  --header @control-plane.headers \
  --data-urlencode 'tenant_id=team-a' \
  | jq 'map({key_id, name, allowed_models, enabled, expires_at})'

curl --fail-with-body --silent --show-error \
  --header @control-plane.headers \
  "$CONTROL_API/config/ai/apikey/$KEY_ID" \
  | jq '{key_id, tenant_id, name, allowed_models, enabled, expires_at}'

Verify that neither response contains raw_key nor key_hash. Avoid listing all tenants unless your role and operational need require it.

Disable, rotate, and delete

Disable is reversible and is useful for a controlled cutover:

curl --fail-with-body --silent --show-error \
  --request PATCH \
  --header @control-plane.headers \
  --header 'Content-Type: application/json' \
  --data '{"enabled":false}' \
  "$CONTROL_API/config/ai/apikey/$OLD_KEY_ID"

Expected result: 204 No Content. The route is described in the companion swagger-extras.yml contract because it is dispatched before the generated OpenAPI handler chain.

1. Create a replacement with the same or narrower permissions.
2. Store it in the workload's secret manager.
3. Roll the workload to the replacement.
4. Verify authorized requests succeed and old-key traffic has stopped.
5. Permanently delete the old key:

curl --fail-with-body --silent --show-error \
  --request DELETE \
  --header @control-plane.headers \
  "$CONTROL_API/config/ai/apikey/$OLD_KEY_ID"

Expected result: 204 No Content; subsequent get returns not found. Prove that inference use fails on the same mode: 4 rule declaring api_key_auth: required. A rule that omits the field does not enter the key gate. A delete is permanent—create a new key if access is needed again.

Disable, allow-list changes, and delete evict the local authentication and key-summary caches before the operation returns. The development HA path also sends a best-effort invalidation to peers. An unreachable or older peer may continue using a cached entry until its five-minute cache TTL expires, so do not describe peer invalidation as an instantaneous cluster-wide revocation guarantee.

Verify data-plane enforcement

Write the inference header without printing its value:

install -m 600 /dev/null ./inference.headers
printf 'X-Api-Key: %s\n' "$INFERENCE_API_KEY" > ./inference.headers

Test	Expected response
Valid key and allowed model	Backend response
No key or unknown key	401 invalid_api_key
Disabled, expired, or revoked key	401
Valid key, disallowed model	403 model_not_allowed
Burst over key or tenant request bucket	429
Required rule with no/evaluable store	503 policy_store_unavailable

Run destructive or throttling probes only with a dedicated non-production tenant. Give each request a non-secret nonce and require backend receipt delta 0 for every 401, 403, 429, and 503; a client status alone does not prove non-delivery. Never print the test key in the result.

Tenant limits

Tenant RPS, aggregate TPM, per-model TPM, and token burst capacity use the tenant rate-limit API. These limits are shared by the tenant's keys and are explained step by step in AI Traffic Governance.

Troubleshooting

Symptom	Likely cause	Action
Create returns 400	Missing tenant or invalid body	Validate JSON and provide a non-empty tenant_id
Management call returns 401	Management credential missing or expired	Refresh through the approved identity workflow
Management call returns 403	Authenticated viewer or unknown role attempted a mutation	Use an explicitly authorized administrator; do not widen the viewer role
Key or quota call returns 503 ai_key_store_unconfigured	--aikey-db-host is unset	Stop exposure, configure the independent key store, and re-run missing-key probes
Key or quota call returns 503 ai_key_store_unavailable	A configured store did not initialize or is unreachable	Restore the store and verify the reconnect; do not bypass the key check
Inference call returns 401	Key missing, unknown, disabled, expired, or revoked	Inspect summary by key_id; do not log the raw key
Required inference call returns 503 policy_store_unavailable	No store, unreachable store without a usable cached answer, or policy evaluation failure	Restore the store; do not relabel this as an invalid client key
Inference call returns 403	Effective model is not allowed	Compare the exact model with allowed_models
Inference call returns 429	Key/tenant RPS or token quota	Inspect the error reason, Retry-After, and metrics
Key cannot be recovered	Raw value was not stored	Revoke the record and rotate to a new key
CRUD fails despite valid management auth	Independent PostgreSQL key store unavailable	Restore the data-plane store securely; management-user health does not prove key-store health

Cleanup

rm -f ./new-key.json ./control-plane.headers ./inference.headers
unset CONTROL_PLANE_TOKEN INFERENCE_API_KEY KEY_ID OLD_KEY_ID

Related pages

• AI Traffic Governance
• SSE and Quota Management
• AI Quotas and QoS
• AI Key Store Operations
• Management API Authentication
• Data-Plane Authentication and JWT
• Monitoring and Metrics