Skip to main content
Voice-agent service KeenableWebSearch gives Pipecat voice agents web search and page fetch as function-call tools, tuned for the latency budget of a spoken conversation. It ships in Pipecat itself. It wraps MCPClient and points it at the hosted Keenable MCP server, so the tools register themselves on the LLM context. → View the change on GitHub

Install

Install the extra

KeenableWebSearch needs Pipecat 1.8.0 or later. Install the extra:

Add it to the context

Build the LLM context with the Keenable tools:
tools() is a coroutine, so call it inside your bot’s async entry point. On first use it connects to the server and returns the tool schemas with their handlers attached, so the LLM registers search_web_pages and fetch_page_content automatically. The connection closes when the pipeline tears down. Call close() only to release it earlier.

Pick a search mode

Keyless calls run in pro mode on the shared public tier, at lower rate limits. Pass an API key to raise the limits and turn on realtime, the lower-latency mode built for voice:
The mode is pinned per call, and the model can’t see or change it. See Choose a search mode to compare the two. If you pass a key and no mode, the default is realtime. If your account doesn’t have realtime enabled, pass mode="pro", or every search fails. → Full example in the Pipecat repo

Tavus

Give a Tavus PAL live web search with one API tool that calls Keenable, so it can answer questions about current events mid-conversation.

Search quickstart

Give your agent ranked web pages with excerpts ready to cite. Try it with no API key. Code, cost per mode, filters by site, date and domain, and gotchas.

Tool calling

Give an LLM web search and page fetch as tools. Pass the SDK’s TOOLS to any OpenAI-compatible chat completions client and run each call with run_tool_call.

Qwen Audio Agent

Give Qwen Audio Agent’s voice assistant Keenable web search with QWEN_AUDIO_WEB_SEARCH_PROVIDER=keenable. No key needed, in release 2.1.0 and later.