One key, one base URL, and you are behind it.

Step 2

Change the base URL

One format per provider, with no translation between the two. The path of your calls does not change: the host changes, and your SDK appends the rest as usual.

Two convenience passthroughs exist as well, with no logging and no budget: GET /v1/models and POST /v1/messages/count_tokens. With several provider keys, they use only the first one.

Step 4

Agent frameworks

An autonomous agent is exactly the traffic you want capped: it loops, it runs away, and nobody is watching. These frameworks all accept a base URL and custom headers, so Outcap plugs in by changing the model configuration.

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="gpt-4o-mini",
    base_url="https://proxy.outcap.tech/v1",
    api_key=os.environ["OUTCAP_KEY"],
    default_headers={
        "x-provider-key": os.environ["OPENAI_API_KEY"],
        "x-outcap-route": "agent-research",
    },
)

from langchain_anthropic import ChatAnthropic

llm = ChatAnthropic(
    model="claude-haiku-4-5",
    base_url="https://proxy.outcap.tech",
    api_key=os.environ["OUTCAP_KEY"],
    default_headers={
        "x-provider-key": os.environ["ANTHROPIC_API_KEY"],
        "x-outcap-route": "agent-research",
    },
)

from crewai import LLM

# CrewAI goes through the OpenAI format, Outcap sits in front
llm = LLM(
    model="openai/gpt-4o-mini",
    base_url="https://proxy.outcap.tech/v1",
    api_key=os.environ["OUTCAP_KEY"],
    extra_headers={
        "x-provider-key": os.environ["OPENAI_API_KEY"],
        "x-outcap-route": "crew-agents",
    },
)
# Agent(..., llm=llm)

Same principle for LlamaIndex, AutoGen or the Vercel AI SDK: any client that exposes a base URL and headers works. Your agents stay in shadow mode until you turn something on.

Reference

Response headers

This is the product's vocabulary. Every decision Outcap makes about a request leaves a readable trace here, streaming included: a model served instead of another, a cap applied, a value redacted, a backup key used.

On an error response from the provider, its own retry headers reach you unchanged: retry-after, retry-after-ms, x-should-retry, the request identifier and the rate limit counters. Its organisation and workspace identifiers do not.

After the first call

Where to start

In this order. The first three change no request at all, and the third is the one that pays.

1. 01Leave it in shadow modeThe default mode changes nothing. A route needs ten answers before a cap is learned, and answers with tool calls are deliberately left out of the calculation.
2. 02Set a hard budgetAn amount per day across the whole project is enough to start. It is the one setting that cannot surprise you: beyond it, the request is refused before the call to the provider, for zero spend.
3. 03Look at the routing suggestionsIn a route's detail view. The price gap is exact, computed on your real tokens and the providers' public prices. It is the only saving in the product that measures to the cent.
4. 04Name your routesSet x-outcap-route on every call. A system prompt fingerprint works, but a table of tags is readable, and per route settings hang off it.
5. 05Pass your customer's identifierIf you bill customers, x-outcap-user unlocks per customer budgets and the margin policy. Without it, Outcap does not know who is spending.