| name | LiteLLM Proxy Authentication Hooks |
| description | How custom_auth, CustomLogger hooks, and FastAPI middleware interact with standard and passthrough endpoints in LiteLLM proxy |
| topics | litellm, proxy, authentication, passthrough, anthropic, custom_auth, hooks |
| created | "2026-04-19T00:00:00.000Z" |
| updated | "2026-04-19T00:00:00.000Z" |
| scratchpad | .specs/scratchpad/516cd93e.md |
LiteLLM Proxy Authentication Hooks
Overview
LiteLLM proxy has two distinct auth hook systems. custom_auth (via general_settings) fires for ALL routes including true HTTP passthrough endpoints. CustomLogger.async_pre_call_hook only fires when LiteLLM routes the request through its SDK — it is silently skipped for true passthrough routes like /anthropic/v1/messages.
Key Concepts
custom_auth: A FastAPI Depends(user_api_key_auth) dependency injected into every route handler — fires before any routing decisionasync_pre_call_hook: A CustomLogger callback that fires inside the LiteLLM SDK call — not triggered for HTTP passthrough- True passthrough: Routes like
/anthropic/{endpoint:path} that forward raw HTTP to the upstream provider — LiteLLM SDK is not involved
- Unified
/v1/messages: A beta endpoint that routes through the LiteLLM SDK (pre_call_hook fires here), distinct from /anthropic/v1/messages
Auth Hook Coverage Matrix
| Hook | /v1/chat/completions | /v1/messages (beta) | /anthropic/v1/messages (passthrough) |
|---|
custom_auth (general_settings) | YES | YES | YES |
async_pre_call_hook (CustomLogger) | YES | YES | NO |
| FastAPI BaseHTTPMiddleware | YES | YES | YES |
Source: verified in litellm/proxy/pass_through_endpoints/llm_passthrough_endpoints.py and litellm/proxy/anthropic_endpoints/endpoints.py on BerriAI/litellm main branch.
Documentation & References
custom_auth: Function Signature and Config
config.yaml
general_settings:
custom_auth: mymodule.my_auth_function
custom_auth_settings:
mode: "on"
Function signature
from fastapi import Request
from litellm.proxy._types import UserAPIKeyAuth
async def my_auth_function(request: Request, api_key: str) -> UserAPIKeyAuth:
token = request.headers.get("authorization", "").removeprefix("Bearer ").strip()
if token != "my-expected-secret":
from litellm.proxy._types import ProxyException
raise ProxyException(message="Unauthorized", type="auth_error", param=None, code=401)
return UserAPIKeyAuth(api_key=api_key, user_id="my-user")
- Must be
async
request: FastAPI Request — access headers via request.headersapi_key: already extracted by LiteLLM from Authorization: Bearer ... or x-api-key header- Return
UserAPIKeyAuth or a str (a valid LiteLLM key to substitute)
- Raise
ProxyException (not a plain Exception) for structured error responses to the client
ProxyException for clean error responses
from litellm.proxy._types import ProxyException
raise ProxyException(
message="Unauthorized: invalid token",
type="auth_error",
param=None,
code=401,
)
How user_api_key_auth Calls custom_auth (Source-Verified)
The flow in litellm/proxy/auth/user_api_key_auth.py:
user_api_key_auth(request, api_key, ...)
└── if enterprise_custom_auth: call it
└── elif user_custom_auth: # ← your custom_auth function
response = await user_custom_auth(request=request, api_key=api_key)
└── else: standard LiteLLM key check
user_custom_auth is populated at startup from general_settings.custom_auth via get_instance_fn().
Passthrough Route Auth (Source-Verified)
/anthropic/{endpoint:path} in llm_passthrough_endpoints.py:
@router.api_route(
"/anthropic/{endpoint:path}",
methods=["GET", "POST", "PUT", "DELETE", "PATCH"],
)
async def anthropic_proxy_route(
endpoint: str,
request: Request,
fastapi_response: Response,
user_api_key_dict: UserAPIKeyAuth = Depends(user_api_key_auth),
):
...
Because user_api_key_auth is a Depends, FastAPI resolves it before the handler runs. custom_auth is called within user_api_key_auth. This is why custom_auth covers passthrough and async_pre_call_hook does not.
Common Pitfalls & Solutions
| Issue | Impact | Solution |
|---|
Using async_pre_call_hook for auth expecting it to cover passthrough | Auth bypass on /anthropic/* | Migrate auth to custom_auth in general_settings |
Raising plain Exception in custom_auth | LiteLLM may return 500 instead of 401 | Raise ProxyException with correct code |
Reading headers from data["proxy_server_request"] in pre_call_hook | Only works in SDK path, not passthrough | In custom_auth, read from request.headers directly |
mode: "auto" when master_key is set | LiteLLM also validates its own keys | Use mode: "on" to rely solely on custom auth |
Custom pass-through endpoints (not built-in) need auth: true | Requests bypass auth | Add auth: true to each entry under general_settings.pass_through_endpoints — Enterprise feature |
Patterns & Best Practices
Migrating from async_pre_call_hook to custom_auth
The existing pattern in request_transformer.py reads headers from data["proxy_server_request"]["headers"] inside async_pre_call_hook. To cover passthrough, the auth check must move to a custom_auth function that receives request: Request directly:
async def async_pre_call_hook(self, user_api_key_dict, cache, data, call_type):
headers = data.get("proxy_server_request", {}).get("headers", {})
token = headers.get("x-litellm-api-key", "")
if token != self.static_token:
raise Exception("Unauthorized")
async def my_auth(request: Request, api_key: str) -> UserAPIKeyAuth:
token = request.headers.get("x-litellm-api-key", "") or api_key
if token != os.environ.get("PROXY_STATIC_TOKEN", ""):
raise ProxyException(message="Unauthorized", type="auth_error", param=None, code=401)
return UserAPIKeyAuth(api_key=token)
Coexistence: custom_auth + CustomLogger for non-auth transforms
Keep CustomLogger (callbacks) for request/response transforms and token rotation. Move only the auth/rejection logic to custom_auth. The two systems are independent and complementary.
Sources & Verification
Changelog
| Date | Changes |
|---|
| 2026-04-19 | Initial creation — LiteLLM custom auth coverage for passthrough endpoints |
