Agent integrations¶

Anything that speaks the OpenAI API works against shaide. In most cases two environment variables are sufficient:

export OPENAI_BASE_URL="https://<endpoint>/v1"
export OPENAI_API_KEY="<key>"

LangChain¶

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(
    model="<model-id>",
    base_url="https://<endpoint>/v1",
    api_key="<key>",
)

LlamaIndex¶

from llama_index.llms.openai_like import OpenAILike

llm = OpenAILike(
    model="<model-id>",
    api_base="https://<endpoint>/v1",
    api_key="<key>",
    is_chat_model=True,
)

Continue (VS Code / JetBrains)¶

{
  "models": [{
    "title": "shaide",
    "provider": "openai",
    "model": "<model-id>",
    "apiBase": "https://<endpoint>/v1",
    "apiKey": "<key>"
  }]
}

Generic OpenAI SDK¶

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://<endpoint>/v1",
  apiKey: "<key>",
});

Running many agents¶

shaide serves multiple models, each with multiple replicas, behind a single endpoint. Agents share that endpoint and the platform load-balances across replicas - no per-agent routing configuration is needed.

To scale capacity, add replicas rather than endpoints. See Scaling.

Troubleshooting¶

Symptom	Cause
`404` on a model	`model` does not match `GET /v1/models`
Client sends to `api.openai.com`	Base URL not applied - some SDKs need `api_base`, not `base_url`
Tool calling ignored	Served model was not trained for tool use
`429` under load	Replicas saturated - add replicas

Extending models with MCP¶

shaide can run Model Context Protocol servers as datasources, giving models access to internal systems — issue trackers, wikis, APIs. Tools exposed by a running datasource become available to models served by the platform.

See MCP servers.