docs: update API.md with service account onboarding guide

- Add roles-based auth for service accounts
- Document client_credentials flow
- Add onboarding steps for new services
- Clarify user vs service auth models

API.md

@@ -633,17 +633,104 @@ Authentik tokens include these custom claims:
 }
 ```
 
-### Capabilities (RBAC)
+### Service Account Auth (Recommended for Services)
 
-The `permissions` claim contains capabilities. Gateway checks these:
+Service accounts use **roles** stored in user attributes. Simpler than groups.
 
-| Capability | Required for | Granted to |
+**JWT claims (client_credentials flow):**
-|------------|--------------|------------|
+```json
-| `llm:inference` | `/v1/chat/completions`, `/v1/embeddings`, `/v1/rerank` | llm-admins, llm-users |
+{
-| `memory:read` | Memory queries | memory-users, memory-writers, poimen-memory-admins |
+  "azp": "portfolio-agent",
-| `memory:write` | Memory ingest | memory-writers, poimen-memory-admins |
+  "roles": ["llm:inference", "memory:read"]
-| `memory:admin` | Memory admin ops | poimen-memory-admins |
+}
-| `*` | Everything | homelab-admins |
+```
+
+**Gateway checks `roles` claim:**
+```go
+if hasRole(claims.Roles, "llm:inference") {
+    // allow
+}
+```
+
+**Available roles:**
+
+| Role | Grants access to |
+|------|------------------|
+| `llm:inference` | `/v1/chat/completions`, `/v1/embeddings`, `/v1/rerank` |
+| `memory:read` | Memory queries |
+| `memory:write` | Memory ingest |
+| `s3:read` | S3/MinIO read |
+| `s3:write` | S3/MinIO write |
+
+---
+
+### Onboarding a New Service
+
+1. **Add to provisioning script** (`homelab/scripts/iam/authentik-provision.py`):
+
+```python
+SERVICE_ACCOUNTS = {
+    # ... existing ...
+    "my-new-service": {
+        "roles": ["llm:inference", "memory:read"],  # what APIs it can call
+        "attributes": {
+            "memory_projects": ["my-project"],  # optional: memory access scope
+            "memory_visibility": "public",
+        },
+        "secret_ns": "my-namespace",       # k8s namespace for credentials
+        "secret_name": "my-service-oidc",  # k8s secret name
+    },
+}
+```
+
+2. **Run provisioning:**
+
+```bash
+export AUTHENTIK_BOOTSTRAP_TOKEN=$(kubectl -n iam get secret authentik-secrets \
+  -o jsonpath='{.data.AUTHENTIK_BOOTSTRAP_TOKEN}' | base64 -d)
+python3 scripts/iam/authentik-provision.py
+```
+
+3. **Use credentials in your service:**
+
+```bash
+# Credentials stored in k8s secret
+kubectl -n my-namespace get secret my-service-oidc -o yaml
+# Contains: CLIENT_ID, CLIENT_SECRET, TOKEN_URL, ISSUER
+
+# Get token
+TOKEN=$(curl -s -X POST $TOKEN_URL \
+  -d "grant_type=client_credentials" \
+  -d "client_id=$CLIENT_ID" \
+  -d "client_secret=$CLIENT_SECRET" \
+  -d "scope=openid roles" | jq -r '.access_token')
+
+# Call API
+curl -H "Authorization: Bearer $TOKEN" https://api.riotpiao.com/v1/models
+```
+
+---
+
+### User Auth (Groups → Permissions)
+
+Human users use **groups** which map to **permissions**. Used for browser-based apps.
+
+**JWT claims (authorization_code flow):**
+```json
+{
+  "groups": ["homelab-admins", "llm-admins"],
+  "permissions": ["*", "llm:inference", "llm:read", "llm:write"]
+}
+```
+
+| Group | Permissions granted |
+|-------|---------------------|
+| `homelab-admins` | `*` (everything) |
+| `llm-admins` | `llm:inference`, `llm:read`, `llm:write` |
+| `llm-users` | `llm:inference` |
+| `memory-users` | `memory:read` |
+| `memory-writers` | `memory:read`, `memory:write` |
+| `poimen-memory-admins` | `memory:*`, `memory:admin` |
 
 **Memory-specific claims:**