AethosHub
PricingRequest Demo

DocumentationAgent Access Manager

LangChain

LangChain

LangChain's OpenAI integrations accept a custom base URL, so the gateway drops in with no framework-specific glue: use ChatOpenAI (or OpenAIEmbeddings) pointed at /v1 with a virtual key, and request models by their catalog alias.

Python

Python
from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    base_url="https://aam.example.com/v1",
    api_key="sk-your-virtual-key",
    model="fast",                 # a catalog alias
)

print(llm.invoke("hello").content)

Embeddings for RAG pipelines route through the same governed pipeline:

Python
from langchain_openai import OpenAIEmbeddings

embeddings = OpenAIEmbeddings(
    base_url="https://aam.example.com/v1",
    api_key="sk-your-virtual-key",
    model="embed",                # an alias whose deployment supports embeddings
)

JavaScript / TypeScript

TS
import { ChatOpenAI } from "@langchain/openai";

const llm = new ChatOpenAI({
  model: "fast",
  apiKey: process.env.AAM_VIRTUAL_KEY,
  configuration: { baseURL: "https://aam.example.com/v1" },
});

LangGraph and agents

Agent workloads are where the gateway earns its keep: an agent loop can burn tokens fast and call tools with sensitive arguments. Everything above applies unchanged to LangGraph — the model client is the same ChatOpenAI — and you get, per agent:

  • One key per agent. Issue each agent its own virtual key so audit, spend, and containment are attributable — quarantining one agent never touches the others.
  • A budget and rate limit on the key as a blast-radius cap for runaway loops: a KEY-scope budget or RPM limit fails the loop with 429 instead of a surprise invoice.
  • Guardrails on tool calls. Tool-call arguments are screened (flag/block, never mutated) — see Guardrails.
  • Vendor swap without redeploys. Re-point the alias in the catalog to move all agents to a new model at once.

Model switching in multi-model chains

Register several aliases (fast, smart, local) and instantiate one ChatOpenAI per alias. Which vendor serves each alias is a catalog decision, not a code decision — including load-balanced and failover configurations.