Intent API Admin

Console-issued admin tokens and zero-boilerplate admin surfaces for any Intent API product.

intent-api-admin plugs into intent-api to add a dedicated admin-machine surface, Ed25519/JWKS-verified admin tokens, an automatic CRUD admin generator from your SQLAlchemy models, role-aware role gates, audit correlation, and a schema document the console can render directly.

Install

pip install intent-api-admin

Requires intent-api>=0.5.0 and Python 3.10+.

Three-step integration

from intent_api import IntentRouter
from intent_api_admin import (
    configure_admin_auth, auto_admin, admin_machine_auth,
)

configure_admin_auth(
    issuer="https://console.example.com",
    audience="myapp",
    base=Base,
)

router = IntentRouter()
router.register("User", UserService())
auto_admin(router, expose=["User", "Team"])

app.include_router(router.build(get_user=get_clerk_user, get_db=get_db))
app.include_router(router.build_admin_machine(get_user=admin_machine_auth(), get_db=get_db))

That is the entire integration. Console operators now get full read/write CRUD on every model in expose=[...], and your handlers are unchanged.

Impersonation (Clerk actor tokens)

Impersonation consumes Clerk's native actor tokens, minted per-product via the product's own Clerk SDK. The console never holds product Clerk credentials and never mints Clerk tokens itself.

The visual indicator is a UX hint, not a guarantee. Always confirm impersonation state via the console's session list before assuming the absence of an indicator means you're not impersonating. The security boundary is server-side.

Two-step integration

Step 1 — Mint. Add a User.create_clerk_actor_token custom action that calls clerk_sdk.actor_tokens.create(...). The full reference adapter is in docs/CLERK_ACTOR_TOKEN_ADAPTER.md.

Step 2 — Enforce. Call enforce_actor_writes at the top of any write handler, OR install actor_writes_middleware once in your runtime.

# Per-handler
from intent_api_admin import enforce_actor_writes

@custom_action(admin_only=True)
async def ban(self, *, db, user, context, id, payload, session_claims=None):
    enforce_actor_writes(session_claims or {}, action="custom",
                         is_read_only_custom_action=False)
    ...

# Or middleware (one-time install)
from intent_api_admin import actor_writes_middleware
runtime = build_runtime(middlewares=[actor_writes_middleware])

Audit context — one line in get_user

bind_impersonation_context populates acting_admin_id and impersonation_session_id as structlog contextvars so every log line in the handler inherits them. One line in your auth dependency:

from intent_api_admin import bind_impersonation_context

async def get_current_user(...):
    auth_state = clerk_sdk.authenticate_request(...)
    bind_impersonation_context(auth_state.payload)   # one-liner
    context.clerk_claims = auth_state.payload
    ...

CSP — frame-ancestors 'none'

The console always launches impersonation in a new tab via window.open(url, "_blank", "noopener,noreferrer"). Set CSP frame-ancestors 'none' on every product response so this architecture self-enforces — even if a future change tries to embed your product, browsers will refuse.

@app.middleware("http")
async def add_csp_headers(request, call_next):
    response = await call_next(request)
    response.headers["Content-Security-Policy"] = "frame-ancestors 'none'"
    return response

See docs/PRODUCT_ADOPTION.md for the full per-product checklist (~60 LOC) and SECURITY.md for the V1 threat model.

Roles

Three roles, ordered by rank:

A higher-ranked role satisfies any lower-ranked requirement. So an owner token can perform any action that requires support, admin, or owner.

Heuristic defaults

auto_admin() reads each SQLAlchemy model and picks sensible defaults you almost never need to override.

• list_fields — primary key + name / email / title / slug / status / created_at if present, then non-text columns up to 6 total.
• search_fields — string columns with index=True or unique=True and length<=255.
• default_sort — created_at desc if present; otherwise <primary_key> desc.
• Auto-redaction patterns — any column whose name starts with password, secret, api_key, token, or _, or ends with _hash / _digest. Redacted fields are stripped from responses and refused on update.

Override patterns

auto_admin(
    router,
    expose=["User", "Team", "Order"],
    redact_fields={"User": ["last_login_ip"]},          # add to auto-redaction
    deny_edit_fields={"User": ["email"]},                # read-only in admin
    deny_delete=["Order"],                                # disallow delete entirely
    overrides={                                           # surgical schema override
        "User": {"default_sort": "email asc",
                 "list_fields": ["id", "email", "name", "plan", "is_active"]},
    },
)

Custom admin actions

Use the augmented @custom_action from intent_api_admin for admin-only commands. They are dispatched on the admin-machine surface only — invisible to standard, MCP, machine, and the legacy admin route.

from intent_api_admin import custom_action

class UserService(IntentService):
    @custom_action(admin_only=True, min_role="admin")
    async def ban(self, *, db, user, context, id, payload):
        ...

    @custom_action(admin_only=True, min_role="support", read_only=True)
    async def view_audit(self, *, db, user, context, id, payload):
        ...

read_only=True is honored by enforce_actor_writes (see above) so the action remains callable under a read-only Clerk actor session.

Common errors and fixes

See TOKEN_CONTRACT.md for the full token spec.

Reference test suite

tests/wbs_reference/ is an exhaustive contract suite that mirrors the WBS success criteria S1–S52. Copy it into your repo to validate every adoption.