Use Augure with LangChain (Python)

You get Augure as the chat model behind LangChain's ChatOpenAI class, so chains, agents, tool binding and streaming run with inference on Canadian infrastructure.

Prerequisites

• Python 3.9+ and pip install langchain-openai.
• An Augure API key (aug_sk_live_…) from the developer dashboard. Plans and pricing: augureai.ca/developers.
• The key exported as AUGURE_API_KEY.

Configure

Use ChatOpenAI from langchain_openai with base_url set to Augure. Set use_responses_api=False so LangChain always sends Chat Completions requests; Augure does not expose a Responses endpoint.

import os
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="tofino-3",
    api_key=os.environ["AUGURE_API_KEY"],
    base_url="https://api.augureai.ca/v1",
    max_tokens=64,
    use_responses_api=False,
)

response = llm.invoke("Say hello in one short sentence.")
print(response.content)

ChatOpenAI also reads OPENAI_API_KEY and OPENAI_BASE_URL from the environment if you leave the arguments out.

Streaming, with usage reported in the last chunk (LangChain sets stream_options.include_usage when stream_usage=True):

llm = ChatOpenAI(model="rosedale-1", base_url="https://api.augureai.ca/v1",
                 api_key=os.environ["AUGURE_API_KEY"], stream_usage=True)
for chunk in llm.stream("Hello"):
    print(chunk.content, end="", flush=True)

Tool calling uses the standard bind_tools path, and structured output works with with_structured_output(Schema, method="function_calling"), since both go through the tools field of the request:

llm_with_tools = llm.bind_tools([search_grant_programs])

Pick a model

Model	When to use
ossington-5	Flagship reasoning. 256K context, 32,768 max output. Default for hard tasks.
rosedale-1	Premium reasoning, up to 1M context on paid plans, Canadian inference. Long documents, agent loops.
tofino-3	Fast everyday model, 256K context, 16,384 max output, no thinking. Chat, extraction, classification.

model="auto" lets the gateway pick. On ossington-5, reasoning_effort="none" turns thinking off and is the fastest setting for mechanical tasks; low, medium and high are accepted on reasoning models; tofino-3 ignores it. ChatOpenAI has a reasoning_effort argument; on older langchain-openai releases pass extra_body={"reasoning_effort": "none"} instead. Full list: /docs/models.

Verify it works

LangChain's response_metadata keeps token usage and the model name but not Augure's extra top-level fields, so check residency with one direct call:

curl -s https://api.augureai.ca/v1/chat/completions \
  -H "Authorization: Bearer $AUGURE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"tofino-3","max_tokens":64,"messages":[{"role":"user","content":"Reply with the single word OK."}]}'

The JSON ends with an _augure block, for example "_augure": {"gateway_region": "ca-montreal-1", "inference_region": "ca-on-toronto", "request_id": "..."}. inference_region tells you where the model ran. Inference is on Canadian infrastructure first, with disclosed failover to vetted EU partners under zero data retention, never the US.

Notes

• message.content holds the final answer only. When a reasoning model thinks, the reasoning text is returned by the API in reasoning_content and is not merged into content.
• For LangGraph agents, pass the same ChatOpenAI instance; nothing else changes.
• OpenAI model aliases are accepted: gpt-4o maps to ossington-4, gpt-4o-mini to tofino-3.
• Use ChatOpenAI, not OpenAI (the legacy completions class); Augure serves /v1/chat/completions.

See also

• API reference
• Models
• MCP server
• IDE setup
• Pricing