Parallel Tool Calls on OpenRouter

What it does

bind_tools on the OpenRouter integration exposes parallel_tool_calls=True as a native option. When enabled, the model can fire multiple tool calls in a single response turn instead of a back-and-forth loop. LangChain handles the streaming chunks internally and merges them into complete tool call sequences.

Why it matters

A research agent that fetches a weather API and a news feed at the same time would otherwise serialize those calls, doubling latency. parallel_tool_calls=True lets the model decide when to act in parallel, and LangChain reassembles the chunks automatically.

Example

$ llm = ChatOpenRouter(model="anthropic/claude-3.5-sonnet", openrouter_api_key="<your-key>", parallel_tool_calls=True)
bound = llm.bind_tools([get_weather, get_news])
resp = bound.invoke("Weather in Vienna and top news in AI?")
for call in resp.tool_calls:
    print(f"  -> {call['name']}({call['args']})")

Tool calls made:
  -> get_weather({'city': 'Vienna'})
  -> get_news({'topic': 'AI'})

Common flags

ChatOpenRouter.bind_tools

Bind tools with optional parallel_tool_calls kwarg

create_openai_functions_chain

Converts tool definitions into an OpenAI-format function-calling chain

create_taggingChain

Structured output via Pydantic schema, an alternative to tool calling