tripwire-litellm

Per-client budget hard stop and cost logging for a LiteLLM proxy. Before each call through your LiteLLM proxy, the plugin checks that client's budget in Tripwire and denies the call with HTTP 402 once the budget is used up. After each call, it logs the cost to Tripwire per client and project, so you can see what each client actually costs you.

Built for AI agencies that run client work through one shared LiteLLM proxy and need to know, and cap, what each client spends.

You need a Tripwire account. There's a free 14-day trial, no card required, at https://www.tripwire.agency. The hard stop applies on the Agency and Scale plans and during an active trial. On Starter (or an expired trial) calls are always allowed and cost is still logged.

Walkthrough video: https://www.youtube.com/watch?v=d6WsQY1OMD4

Install

Python 3.9 or newer, on the machine that runs your LiteLLM proxy (LiteLLM proxy 1.65.0 or newer).

pip install tripwire-litellm

Config

Point callbacks at the handler instance in your proxy config:

litellm_settings:
  callbacks: ["tripwire_litellm.handler.proxy_handler_instance"]

Then set the environment variables (defaults shown):

TRIPWIRE_INGEST_TOKEN=tw_live_...          # required, never log it
TRIPWIRE_BASE_URL=https://www.tripwire.agency
TRIPWIRE_FAIL_MODE=open                    # or closed
TRIPWIRE_CHECK_TIMEOUT_MS=800
TRIPWIRE_DENY_STATUS=402                   # not 429; SDKs retry 429
TRIPWIRE_RESERVE_USD=0
TRIPWIRE_HARD_STOP=on                      # off = ingest only
TRIPWIRE_CLIENT_ID=
TRIPWIRE_PROJECT_ID=

Calls are attributed to a client and project from, in order: request metadata (tripwire_client_id, tripwire_project_id), the virtual key's metadata, the team's metadata, then TRIPWIRE_CLIENT_ID / TRIPWIRE_PROJECT_ID.

Limits

• Fail-open by default. If Tripwire can't be reached (timeout, network error, 5xx, 401), the call is allowed. Set TRIPWIRE_FAIL_MODE=closed to return 503 instead. After 5 failures in a row, checks are skipped for 30 seconds.
• Spend is known after the call. Tripwire learns the cost when a call finishes and is ingested, so calls already running when the budget is crossed still complete. Worst-case overshoot is about 2 × concurrency × average call cost, and never less than your single largest call.
• Fixed reserve. TRIPWIRE_RESERVE_USD sends the same estimated cost with every check. There is no per-request sizing from tokens. Lower concurrency on a key, a reserve, or fail-closed mode reduce overshoot.

For developers

Hooks (checked against the LiteLLM proxy call-hook docs, 2026-10-05):

• async_pre_call_hook(self, user_api_key_dict, cache, data, call_type): returns data unchanged on allow; on deny raises fastapi.HTTPException 402 with error.code tripwire_budget_exceeded.
• async_log_success_event(self, kwargs, response_obj, start_time, end_time): call id from kwargs["litellm_call_id"], then kwargs["litellm_params"]["litellm_call_id"], then response_obj.id; cost and provider from kwargs["response_cost"], kwargs["litellm_params"]["custom_llm_provider"], kwargs["model"].

Unit tests stub litellm.integrations.custom_logger when LiteLLM isn't installed.