Drop-in proxy

The proxy records each AI call and forwards it to the real provider, so you get request-level attribution without touching application logic. You change one thing: the provider SDK’s base URL. The <your-venturi-instance> placeholder below is your data-plane host in either deployment mode (see Deployment modes); your onboarding contact provides it.

Fail-open guarantee

The proxy runs on a hard latency budget. If Venturi is slow or unreachable, the request is forwarded to the provider anyway. The proxy can never block or fail your production traffic.

How it works

The proxy passes your request through unchanged (including your provider API key, which Venturi does not store), captures metadata (model, tokens, latency, cost) and emits an InvocationEvent. Request and response content exists only in memory while being forwarded and is never written to access logs, error logs, traces, queues, or exception payloads.

Endpoints

Relative to your Venturi instance host (provided during onboarding):

Provider	Proxy base path
OpenAI	/api/v1/proxy/openai/v1
Anthropic	/api/v1/proxy/anthropic/v1
Amazon Bedrock	/api/v1/proxy/bedrock

OpenAI example

PythoncURL

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://<your-venturi-instance>/api/v1/proxy/openai/v1",
    api_key="<your OpenAI key>",   # forwarded as-is; not stored by Venturi
)

client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "…"}],
)

Bash

curl https://<your-venturi-instance>/api/v1/proxy/openai/v1/chat/completions \
  -H "Authorization: Bearer <your OpenAI key>" \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-4o","messages":[{"role":"user","content":"…"}]}'

Anthropic example

PythoncURL

Python

from anthropic import Anthropic

client = Anthropic(
    base_url="https://<your-venturi-instance>/api/v1/proxy/anthropic/v1",
    api_key="<your Anthropic key>",
)

client.messages.create(
    model="claude-sonnet-4-6",
    max_tokens=1024,
    messages=[{"role": "user", "content": "…"}],
)

Bash

curl https://<your-venturi-instance>/api/v1/proxy/anthropic/v1/messages \
  -H "x-api-key: <your Anthropic key>" \
  -H "anthropic-version: 2023-06-01" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-sonnet-4-6","max_tokens":1024,"messages":[{"role":"user","content":"…"}]}'

Attributing the call

To attribute beyond “this API key”, forward identity/service hints as headers your platform team configures (e.g. an identity header or your existing trace headers). Ask your onboarding contact for the header convention enabled on your instance.

Streaming

OpenAI and Anthropic Server-Sent Events are forwarded incrementally and token counts are derived from the stream. Bedrock non-streaming requests use the route above. Bedrock binary event streams are forwarded but are not parsed for token usage; those rows are labeled estimated rather than presented as complete request-level metering.

The decision-time interceptor that supplies routing guidance has a 50 ms P99 end-to-end budget; proxy forwarding remains fail-open if that budget is exhausted.

Confirm events are landing