latchvector-sso

Python SDK for Latch Vector SSO. Python 3.9+.

pip install latchvector-sso

Framework integrations live in submodules and pull nothing in unless you import them:

pip install "latchvector-sso[fastapi]"   # or [flask], [django]

---

Contents

• Protecting an API — the common case
• What audience is for
• The principal
• Logging users in
• Password reset
• Device sessions (mobile)
• Machine-to-machine (API clients)
• Multitenancy
• Errors
• Configuration
• Smoke test
• Before you go live
• Management API
• Webhooks
• Migrating from your current system

Protecting an API — the common case

Most integrations only need this. Your API verifies tokens locally; it does not call the SSO service on every request.

from latchvector_sso import TokenVerifier

# Build once at startup — it caches the discovery document and signing keys.
verifier = TokenVerifier(
    issuer="https://sso.yourdomain.com",
    audience="https://api.yourcompany.com",   # your registered identifier
)

principal = verifier.verify_authorization_header(request.headers.get("Authorization"))

FastAPI

from fastapi import Depends, FastAPI
from latchvector_sso import Principal, TokenVerifier
from latchvector_sso.fastapi import SsoAuth

auth = SsoAuth(TokenVerifier(issuer=..., audience=...))
app = FastAPI()

@app.get("/invoices")
def list_invoices(user: Principal = Depends(auth.required)):
    return {"owner": user.uid}

@app.post("/invoices/{invoice_id}/approve")
def approve(invoice_id: int, user: Principal = Depends(auth.requires("invoice.approve"))):
    ...

Flask

from latchvector_sso.flask import SsoAuth, current_principal

auth = SsoAuth(app, TokenVerifier(issuer=..., audience=...))

@app.get("/invoices")
@auth.required
def list_invoices():
    return {"owner": current_principal().uid}

@app.post("/invoices/<int:invoice_id>/approve")
@auth.requires("invoice.approve")
def approve(invoice_id):
    ...

Django

# settings.py
LATCHVECTOR_SSO = {
    "ISSUER": "https://sso.yourdomain.com",
    "AUDIENCE": "https://api.yourcompany.com",
}
MIDDLEWARE = [..., "latchvector_sso.django.SsoAuthenticationMiddleware"]

# views.py
from latchvector_sso.django import sso_required, sso_requires

@sso_required
def invoices(request):
    return JsonResponse({"owner": request.principal.uid})

@sso_requires("invoice.approve")
def approve(request, pk):
    ...

The middleware attaches request.principal when a valid token is present and leaves it None otherwise; the decorators do the rejecting. That keeps authentication and authorisation separable, so a view can be deliberately public without having to bypass the middleware.

---

What audience is for

It is your application's registered identifier, and it is required — there is no option to turn the check off.

A token issued for a different application is still validly signed by a trusted issuer. If you check the signature but not the audience, you accept it, which means you accept one from every user of every application on the platform. This is the single most common way an SSO integration is compromised, and it is why the parameter has no default.

The principal

principal.uid          # 4711 — key your records on this
principal.email        # display only, see below
principal.org_id       # 57
principal.tenant_id    # 1
principal.org_path     # "/1/57/"
principal.permissions  # frozenset({"invoice.approve"})
principal.expires_at   # timezone-aware datetime

principal.has("invoice.approve")
principal.has_any("invoice.approve", "invoice.admin")
principal.has_all("invoice.read", "invoice.approve")
principal.can_reach("/1/57/903/")   # does their granted scope cover this node?

Key your own tables on uid, never on the email. Addresses change, and a GDPR erasure request scrubs the address while uid survives. Rows keyed on email lose the link to their own user the first time either happens.

Do not cache permissions past expires_at. They are a snapshot from issue time; a revoked role takes effect on the next token, which is why access tokens last only 15 minutes.

---

Logging users in

Only whatever actually handles the password needs this — a login backend, a BFF, a mobile gateway. Your resource APIs do not.

from latchvector_sso import SsoClient, MfaRequired

sso = SsoClient(
    issuer="https://sso.yourdomain.com",
    audience="https://api.yourcompany.com",
)

result = sso.login(email, password)

if isinstance(result, MfaRequired):
    code = prompt_user_for_code()               # TOTP or recovery code
    tokens = sso.verify_mfa(result.pending_token, code)
else:
    tokens = result

login() returns a union, not an object with an empty access_token. You have to branch before you can reach a token, so the MFA path cannot be quietly skipped — a customer will enable MFA eventually, and the failure mode of the nullable version is an empty token in production.

Social login

result = sso.social_login("google", google_id_token)

The user must already exist. Accounts are provisioned by an administrator; a first-time social login for an unknown email is refused rather than silently creating an account.

Refresh

fresh = sso.refresh(stored_refresh_token)
save_refresh_token(fresh.refresh_token)   # before you use it

Refresh tokens rotate: the old one is dead the moment refresh() returns. Persist the new one first.

from latchvector_sso import RefreshTokenError, RefreshTokenReusedError

try:
    return sso.refresh(stored)
except RefreshTokenReusedError:
    destroy_session()
    alert_security_team(user_id)      # this is a security event
    raise
except RefreshTokenError:
    return redirect_to_login()        # expired or unknown — ordinary

RefreshTokenReusedError is deliberately not a subclass of RefreshTokenError, so an except that only meant "refresh or re-login" cannot swallow a compromise signal.

Logout

sso.logout(refresh_token)

This revokes the refresh token. The current access token stays valid for the rest of its 15 minutes — it is a signed bearer token, not a session. For immediate cut-off, have an administrator disable the account.

---

Machine-to-machine (API clients)

For a backend job that acts as itself, not a user — the OAuth2 client_credentials grant. An admin registers an API client (secret shown once) bound to an application; the job exchanges the credentials for a short-lived token.

Each framework has a machine counterpart of the user helpers. They verify with verify_client, so a user access token is rejected there just as a machine token is rejected by the user helpers — the two never cross.

FastAPI

from latchvector_sso import ClientPrincipal
from latchvector_sso.fastapi import SsoAuth

auth = SsoAuth(TokenVerifier(issuer=..., audience=...))

@app.post("/reports/sync")
def sync(client: ClientPrincipal = Depends(auth.requires_scope("reports.write"))):
    return {"org": client.org_id, "client": client.client_id}

Flask

from latchvector_sso.flask import SsoAuth, current_client

@app.post("/reports/sync")
@auth.requires_scope("reports.write")
def sync():
    return {"org": current_client().org_id}

Django

from latchvector_sso.django import sso_requires_scope

@sso_requires_scope("reports.write")
def sync(request):
    return JsonResponse({"org": request.client.org_id})

Calling another service (your app is the job) — obtain and cache a token:

machine = sso.client_credentials(client_id, client_secret, ["reports.write"])
httpx.post(url, headers={"authorization": f"Bearer {machine.access_token}"})
# machine.expires_in_seconds ~ 900; no refresh — cache and re-fetch on expiry.

---

Multitenancy

Verifying a token tells you who is calling; multitenancy is about what data they may touch. The SDK ties your models to the tenant in the verified token, so a query cannot read or write another tenant's rows even if you forget the filter. It works on either ORM — Django's, or SQLAlchemy for Flask and FastAPI.

Django

Inherit BelongsToTenant (it adds a tenant_id column and the scoping):

from latchvector_sso.django import BelongsToTenant

class Invoice(BelongsToTenant):
    amount = models.IntegerField()

Behind SsoAuthenticationMiddleware, that is all:

Invoice.objects.all()            # only the caller's tenant
Invoice.objects.create(amount=5) # tenant_id stamped automatically
Invoice.all_tenants.all()        # escape hatch: every tenant, used deliberately

Configure it in settings.py:

LATCHVECTOR_SSO = {
    "ISSUER": "...",
    "AUDIENCE": "...",
    "TENANT": {
        "ENABLED": True,                      # False in a sandbox / dev
        "COLUMN": "tenant_id",
        "BYPASS_PERMISSIONS": ["PLATFORM_ADMIN"],  # see across tenants
    },
}

• Bypass — a caller with a BYPASS_PERMISSIONS code (a platform operator) is unconstrained; an org admin is still bound to their tenant.
• Sandbox — set "ENABLED": False so dev data and tests aren't confined.
• Commands & tasks — with no request there is no tenant, so the scope is inert. In a task that must be tenant-bound, set it: TenantContext.set(id).

The TenantContext primitive is importable directly (from latchvector_sso import TenantContext) if you integrate a different ORM.

Flask & FastAPI (SQLAlchemy)

Those frameworks use SQLAlchemy, so the model layer is a mixin plus one call at startup. The SSO auth (below) sets the tenant context from the verified token; the rest is identical to Django.

from latchvector_sso.sqlalchemy import BelongsToTenant, install_tenant_scoping

class Invoice(Base, BelongsToTenant):          # gains a tenant_id column
    __tablename__ = "invoices"
    id = mapped_column(Integer, primary_key=True)
    amount = mapped_column(Numeric)

install_tenant_scoping()                       # once, after your models import

That is all — reads gain WHERE tenant_id = <current> (through joins and relationship loads too) and inserts are stamped:

session.scalars(select(Invoice)).all()   # only the caller's tenant
session.add(Invoice(amount=5))            # tenant_id stamped on flush

The auth integrations set the context automatically once you tell them who may see across tenants:

# FastAPI
auth = SsoAuth(verifier, bypass_permissions=["PLATFORM_ADMIN"])
# Flask
auth = SsoAuth(app, verifier, bypass_permissions=["PLATFORM_ADMIN"])

• Bypass — a caller holding a bypass permission is unconstrained; an org admin stays bound to their tenant.
• Async-safe — the context lives in a ContextVar, correct under both thread-per-request (Flask) and many concurrent requests on one event loop (async FastAPI).
• Sandbox / other sessions — install_tenant_scoping(enabled=False) turns scoping off; pass your own Session/sessionmaker class as the session argument if you don't use the default.

Confining to a sub-tree

tenant_id is the hard wall between customers. Within one customer, an admin of a sub-org should often see only their slice of the org tree, not the whole tenant. Opt a model into subtree mode and it is narrowed to exactly the org paths the caller's token grants:

# Django
class Chart(BelongsToTenant):
    tenant_scope_mode = "subtree"          # default is "tenant"
    org_id = models.BigIntegerField(editable=False, null=True)
    org_path = models.TextField(editable=False, null=True)  # index it: see below
    title = models.CharField(max_length=120)

# SQLAlchemy (Flask / FastAPI) — a ready-made mixin with both columns
from latchvector_sso.sqlalchemy import BelongsToTenantSubtree

class Chart(Base, BelongsToTenantSubtree):   # adds org_id + org_path
    __tablename__ = "charts"
    id = mapped_column(Integer, primary_key=True)
    title = mapped_column(String)

Alongside tenant_id, a subtree model needs an org_id and an org_path column (a materialized path like /1/57/903/). New rows are stamped with the writer's own node; reads are confined to:

• SUBTREE grants — the caller's node and everything below it (a left-anchored org_path__startswith);
• SELF grants — that node only (an exact match).

Which applies is decided by the caller's roles at token-issue time and carried in the scope_subtree / scope_self claims — you write nothing. A machine (client-credentials) token has no org reach, so a subtree model falls back to tenant-wide for it — still leak-safe across customers.

The trailing slash matters. Paths are stored /1/57/ (not /1/57), so the prefix /1/57/ can never leak into a sibling like /1/570/.

Multitenancy at scale

For tables that will hold billions of rows, three columns and the right indexes keep every scoped query a range scan, never a table scan:

Column	Type	Why
tenant_id	bigint	the hard customer wall; on every tenant-aware table
org_id	bigint	the owning node — subtree tables only
org_path	text	materialized path /1/57/903/, trailing slash — subtree only

Index tenant-leading, so the tenant predicate drives the scan:

-- every tenant-aware table
CREATE INDEX ON invoices (tenant_id, created_at DESC);

-- subtree tables: prefix scans on org_path within the tenant
CREATE INDEX ON charts (tenant_id, org_path text_pattern_ops);

text_pattern_ops is what makes org_path LIKE '/1/57/%' an index range scan under any collation. For the largest tenants, partition or shard by tenant_id (Postgres declarative partitioning, or Citus/Nile-style distribution): the tenant-leading key means a query already touches only its own partition.

---

Errors

Every error is an SsoError with .code and .status.

Class	Codes
AuthenticationError	invalid_credentials, invalid_code, invalid_id_token, invalid_token, invalid_token_use, invalid_or_expired_pending_token
RefreshTokenError	invalid_refresh_token, refresh_token_expired
RefreshTokenReusedError	refresh_token_reused
AccountNotActiveError	account_not_active
AccountLockedError	account_locked
AccessDeniedError	access_denied
ValidationError	validation_failed (with .fields)
RateLimitError	too_many_requests (with .retry_after_seconds)
ConfigurationError	unknown_audience, discovery failures

429 is retried automatically with exponential backoff and jitter (twice by default, max_rate_limit_retries to change it). Nothing else is retried, and error.retryable is False for everything but RateLimitError. A 403 is a decision the service already made; retrying it produces a stream of ACCESS_DENIED audit entries that a compliance officer will eventually ask you about.

access_denied does not distinguish "forbidden" from "does not exist" — telling them apart would let anyone enumerate records across tenants.

---

Configuration

You configure one URL. The JWKS endpoint is resolved from {issuer}/.well-known/openid-configuration and cached, so the SDK keeps working if it ever moves.

Argument	Default
issuer	—	required
audience	—	required, cannot be disabled
leeway_seconds	30	skew allowance; keep NTP running regardless
jwks_cache_seconds	600
timeout_seconds	10.0
max_rate_limit_retries	2	client only

TokenVerifier is thread-safe and intended to be shared.

---

Smoke test

SSO_ISSUER=http://localhost:9000 SSO_AUDIENCE=http://localhost:9000 \
SSO_EMAIL=… SSO_PASSWORD=… python examples/smoke.py

Beyond the happy path it asserts that a token minted for a different audience is rejected and that a tampered signature is rejected — the two checks whose absence turns a working integration into an open door.

---

Before you go live

• audience is set to your identifier, not ours
• Your tables key on uid, not email
• TokenVerifier is constructed once, not per request
• RefreshTokenReusedError is handled as a compromise, not retried
• The MfaRequired branch is implemented and tested
• Tokens are never written to logs, URLs, or error reports

Password reset

The invite / forgot-password flow (the token comes from the emailed link or an admin-issued setup link):

sso.forgot_password(email)                # emails a one-time link (no account oracle)
sso.reset_password(token, new_password)   # redeem the link's token

Device sessions (mobile)

A mobile app gets a longer-lived, device-bound session by passing a device at login (also on verifyMfa/socialLogin). The service returns a stable deviceId — store it in secure storage and resend it so the same device is reused. The refresh token lives far longer than the web one and slides on every use; the user can list and revoke devices via GET/DELETE /api/users/me/devices (also on the ManagementClient).

from latchvector_sso import DeviceInfo
# Presence of a device ⇒ a longer-lived, device-bound session.
r = sso.login(email, password, DeviceInfo(name="Ana's iPhone", platform="ios"))
save_to_secure_store(r.device_id)   # store it; resend as DeviceInfo(device_id=…) next launch

Management API

Everything the console does, in code — users, organizations, roles, applications, API clients, webhooks, audit, bulk import, GDPR. Authenticated with a management token (log in with audience equal to the issuer):

from latchvector_sso import SsoClient, ManagementClient

sso = SsoClient(issuer=issuer, audience=issuer)      # management token
tokens = sso.login(email, password)
mgmt = ManagementClient(issuer, tokens.access_token)
mgmt.users.create(organizationId=org_id, email=email, fullName=name)
mgmt.request("POST", "/api/anything", body={"x": 1})  # every endpoint, incl. new ones

→ Management API guide — every resource, the token model, and the generic request() escape hatch.

Webhooks

Get notified the moment a user's access changes — a role assigned or revoked, a role's permissions changed, an account disabled or erased — so you can clear caches or force a refresh instead of waiting for the next failed call. Every delivery is HMAC-signed and timestamped, and this SDK ships a one-call verifier.

→ Webhooks guide — events, payload, the signature scheme, and a verified handler example.

Migrating from your current system

Bring an existing estate — organizations, users, roles, permissions — across in one validated pass. Records reference each other by your own ids, bcrypt passwords carry over (everyone else is invited), and re-runs are safe.

→ Migration guide — the two-step validate/commit flow, the full payload schema, and a worked example.