LangChain

LangChain integrates with Curate-Me through the base_url parameter on ChatOpenAI. No middleware, no monkeypatching — the SDK’s existing configuration surface is enough.

Minimal example

import os
from langchain_openai import ChatOpenAI
 
llm = ChatOpenAI(
    base_url=os.environ.get("CM_GATEWAY_URL", "https://api.curate-me.ai/v1/openai"),
    api_key="unused",
    model="gpt-4o",
    default_headers={"X-CM-API-Key": os.environ["CM_API_KEY"]},
)
 
print(llm.invoke("What is the Curate-Me gateway?").content)

The full example lives at examples/integrations/langchain/main.py. Its smoke test (tests/test_smoke.py) runs on every PR.

What you get for free

• Cost tracking per call (token-accurate, billed in USD)
• Rate limiting via IETF RateLimit-* response headers
• PII scanning on prompts before they reach the upstream provider
• Model allowlist enforcement
• HITL approval gates for high-cost calls
• Full request trace in the Curate-Me dashboard

Supported versions

Known limitations

• Streaming: llm.stream(...) works, and per-chunk content is proxied unmodified. The final cost-record event lands when the stream closes — X-CM-Request-Id is in the headers (sent before the stream body) so you can correlate before the stream finishes.
• Tools: function-calling works via the standard bind_tools(...) API. The gateway tracks tool-call cost as part of the parent request, not as a separate trace row.
• Embeddings: use OpenAIEmbeddings the same way; route through /v1/openai/embeddings. Verified by the mock gateway in CI.

Troubleshooting

# Sanity check: make a call and inspect the response headers.
from openai import OpenAI
client = OpenAI(
    base_url="https://api.curate-me.ai/v1/openai",
    api_key="cm_sk_test_key",  # placeholder — real key from dashboard
    default_headers={"X-CM-API-Key": os.environ["CM_API_KEY"]},
)
resp = client.with_raw_response.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "ping"}],
)
print("request_id:", resp.headers.get("X-CM-Request-Id"))

If X-CM-Request-Id is missing, you’re talking directly to OpenAI — re-check the base_url.