Skip to main content

AI Gateway management API reference

API docs by Redocly

Stacklok LLM Gateway Management API (0.1.0)

Download OpenAPI specification:Download

URL: https://github.com/stacklok/stacklok-llm-gateway License: Apache-2.0

Admin-scoped HTTP API backing the help-desk UI. The server is a thin facade over the Kubernetes API: write endpoints translate to typed AIPolicy CR create/update/delete calls via controller-runtime, and introspection endpoints read from the same cache. No parallel store.

Scope

Each API server instance is bound to exactly one Kubernetes namespace (configured via operator flag) and all AIPolicy CRs live in that namespace. The namespace is therefore not in the URL path; multi- namespace deployments run multiple API instances. A future major version may introduce /v1/namespaces/{ns}/... if multi-tenant scoping is required; clients should treat the current paths as namespace-local.

Authentication and authorization

All requests require a bearer JWT validated against the OIDC provider configured on the target AIGateway (spec.auth.oidc). Role checks are performed per path; role names (admin, policyEditor, authenticated) are conventional and are themselves resolved against AIGateway.spec.auth.authz.roles, which maps each role name to a list of principal matchers.

Concurrency control

GET on a single policy returns an ETag header whose value is the quoted metadata.resourceVersion. PATCH and DELETE both require the If-Match header. Concurrency failures use distinct status codes so the UI can branch without parsing error strings:

  • 428 Precondition Required (RFC 6585) — If-Match header is missing. This is a client bug; UIs should surface it as a dev-tools error rather than a user-facing "policy changed" dialog.
  • 412 Precondition Failed (RFC 7232) — If-Match was sent but the value no longer matches the server's resourceVersion (another admin edited concurrently). UIs should re-fetch and offer a merge/overwrite flow.
  • 409 Conflict — reserved for POST with a metadata.name that already exists. UIs should prompt for a different name.

Catalog

List all MCP servers visible to the gateway.

Forward-compatible; always returns an empty items array in Phase 8. Enforcement lands in a later phase.

Authorizations:
BearerAuth
header Parameters
X-Request-Id
string

Correlation ID; echoed on responses and included in audit records.

Responses

Response samples

Content type
application/json
{
  • "items": [
    ]
}

List all models visible to the gateway in the current namespace.

Derived from the owning AIGateway's spec.routes[].match.model. Deduplicated and sorted by model id, OpenAI-shaped (object:"list", data[]) to match the data-plane GET /v1/models intercept.

Authorizations:
BearerAuth
header Parameters
X-Request-Id
string

Correlation ID; echoed on responses and included in audit records.

Responses

Response samples

Content type
application/json
{
  • "data": [
    ],
  • "object": "list"
}

Return the owning gateway's identity.

Returns the AIGateway name and namespace this api-key-service serves. Sourced from the operator-injected GATEWAY_NAME / GATEWAY_NAMESPACE — authoritative and unambiguous.

Authorizations:
BearerAuth
header Parameters
X-Request-Id
string

Correlation ID; echoed on responses and included in audit records.

Responses

Response samples

Content type
application/json
{
  • "name": "prod-gateway",
  • "namespace": "llm-gateway"
}

Introspection

Return the caller's identity and resolved policy bindings.

Returns subject, groups, resolved roles, the set of AIPolicies whose principalMatchers the caller satisfies, and the union of models granted by those policies' budgets. Matched-only — the response never includes policies the caller does not match.

Authorizations:
BearerAuth
header Parameters
X-Request-Id
string

Correlation ID; echoed on responses and included in audit records.

Responses

Response samples

Content type
application/json
{
  • "email": "admin@example.com",
  • "groups": [
    ],
  • "matchedPolicies": [
    ],
  • "models": [
    ],
  • "roles": [
    ],
  • "subject": "github|42",
  • "unrestrictedModels": false
}

Policies

Create a new AIPolicy.

Use metadata.name as the idempotency key. Retrying a POST with the same name returns 409 Conflict; clients should treat this as either a duplicate submission (succeeded previously) or a deliberate collision (surface to the user).

Authorizations:
BearerAuth
header Parameters
X-Request-Id
string

Correlation ID; echoed on responses and included in audit records.

Request Body schema: application/json
required

Full AIPolicy object. metadata.namespace is ignored; the server's configured namespace wins.

One of
object

Responses

Request samples

Content type
application/json
Example
{ }

Response samples

Content type
application/json
{
  • "apiVersion": "string",
  • "kind": "string",
  • "metadata": {
    },
  • "spec": {
    },
  • "status": {
    }
}

Delete an AIPolicy.

If-Match is required; missing returns 428 and stale returns 412, so accidental wipes during a concurrent edit never silently succeed.

Authorizations:
BearerAuth
path Parameters
name
required
string

Policy name.

header Parameters
If-Match
required
string

Quoted resourceVersion from a preceding GET. Required.

X-Request-Id
string

Correlation ID; echoed on responses and included in audit records.

Responses

Response samples

Content type
application/json
{
  • "error": "unauthenticated",
  • "requestId": "e7c61d23-0b1a-4d3e-8a0b-18d4f1d7e2c3",
  • "violations": [
    ]
}

Get a single AIPolicy by name.

Authorizations:
BearerAuth
path Parameters
name
required
string

Policy name (DNS-1123 subdomain, max 253 chars).

header Parameters
X-Request-Id
string

Correlation ID; echoed on responses and included in audit records.

Responses

Response samples

Content type
application/json
{
  • "apiVersion": "string",
  • "kind": "string",
  • "metadata": {
    },
  • "spec": {
    },
  • "status": {
    }
}

List AIPolicies in the configured namespace.

Returns a page of AIPolicies. limit caps the underlying apiserver page size (max 500; larger values are clamped). pageToken resumes a prior list at the next page boundary. labelSelector accepts the standard apimachinery selector syntax (e.g., team=eng,tier!=internal). gatewayRef is a post-list filter applied to the returned page; when used with limit, the returned page may contain fewer matches than the limit — callers must keep paging while nextPageToken is non-empty to enumerate every match. When labelSelector is set, the apiserver omits remainingItemCount (per the Kubernetes API spec: the remaining count is unknown for selector-filtered lists), so the field will be absent on the response regardless of how many pages remain. A present-but-empty value for limit, pageToken, or labelSelector (e.g., ?limit=) is equivalent to the parameter being absent and uses the server default for that parameter — empty is not "explicitly clear this filter".

Authorizations:
BearerAuth
query Parameters
limit
integer

Max items per page (1..500). Values above 500 are clamped to 500; 0 or omitted uses the server default (500).

pageToken
string

Opaque continue token from a previous response. An expired or malformed token returns 400; restart the list from page one.

labelSelector
string

Kubernetes label selector (e.g., team=eng,tier!=internal). Malformed selectors return 400.

gatewayRef
string

Restrict to policies targeting this AIGateway name. Applied as a post-list filter; may reduce the returned page size below limit.

header Parameters
X-Request-Id
string

Correlation ID; echoed on responses and included in audit records.

Responses

Response samples

Content type
application/json
{
  • "apiVersion": "string",
  • "items": [
    ],
  • "kind": "string",
  • "nextPageToken": "string",
  • "remainingItemCount": 0
}

Update an AIPolicy using JSON merge patch.

Patches are applied against spec only; any status or metadata.resourceVersion sent in the body is ignored (use If-Match for optimistic concurrency). spec.gatewayRef is immutable after creation; attempting to modify it returns 422.

Authorizations:
BearerAuth
path Parameters
name
required
string

Policy name.

header Parameters
If-Match
required
string

Quoted resourceVersion from a preceding GET. Required: missing returns 428 Precondition Required; stale returns 412 Precondition Failed.

X-Request-Id
string

Correlation ID; echoed on responses and included in audit records.

Request Body schema:
required

RFC 7396 JSON merge patch; only /spec may be patched.

object (AIPolicySpecPatch)
description
string <= 2048 characters
object (MCPPolicy)

MCPPolicy binds matched principals to MCP servers. Forward-compatible; not enforced in Phase 8.

Array of objects (PrincipalMatcher)

When present, replaces the stored principalMatchers entirely. 1..32 entries.

Responses

Request samples

Content type
{
  • "spec": {
    }
}

Response samples

Content type
application/json
{
  • "apiVersion": "string",
  • "kind": "string",
  • "metadata": {
    },
  • "spec": {
    },
  • "status": {
    }
}