LlamaIndex

OpenAILike takes api_base. Note the spelling — it is not base_url. The class ships in its own package and is a thin wrapper over the OpenAI model, made compatible with third-party endpoints.

pip install llama-index-llms-openai-like

from llama_index.llms.openai_like import OpenAILike

llm = OpenAILike(
    model="zen5",
    api_base="https://api.hanzo.ai/v1",
    api_key="sk-...",
    context_window=1000000,
    is_chat_model=True,
    is_function_calling_model=True,
)

print(llm.complete("Hello"))

Lands on POST /v1/chat/completions.

Three defaults have to be overridden or the wrapper is wrong about the model. is_chat_model defaults to False, which sends complete() to the completions endpoint rather than chat. is_function_calling_model defaults to False, so an agent built on the model will produce prose where you expected a tool call. And context_window defaults to OpenAI's, not the model's — take the real one from that model's record in GET /v1/models.

llama-index-embeddings-openai-like is the matching package for the retrieval side, and reaches POST /v1/embeddings.

MCP

BasicMCPClient takes the server URL as its first positional argument and infers the transport from it.

pip install llama-index-tools-mcp

from llama_index.tools.mcp import BasicMCPClient, McpToolSpec

client = BasicMCPClient("https://api.hanzo.ai/v1/mcp")
tools = await McpToolSpec(client=client).to_tool_list_async()

A URL ending /mcp is streamable HTTP, one ending /sse is SSE, and a first argument that is a command with args is a local process.