Skip to content

Pipecat

Preview · register_function per spec · pip install "agent-adapter[pipecat]" (pipecat-ai 1.13 to below 2)

Contract tests run against the real pipecat-ai 1.16.0 classes, and interrupt_bench_frameworks.py drives a real pipeline's function calls and interruptions (0% half-done bookings, against 84% on Pipecat's defaults). voice_live_pipecat.py runs voice_processor() with real local audio, local speech-to-text and text-to-speech, and a scripted caller who cuts in mid-booking: charged once in every run (results). No person has run it by voice yet.

Two ways in:

  • Your tools on a Pipecat LLM service. Pipecat's LLM decides; each call runs through the executor.
  • A VoiceAgent in the pipeline. It takes the LLM's place: stt -> processor -> tts, with the voice agent's cut rules and spoken approvals.

Functions

Function (agent_adapter.integrations.pipecat) Returns
pipecat_tools(llm, registry, *, executor=None) registers every tool exported to pipecat on llm; returns their ToolsSchema
pipecat_tool(spec, *, executor=None) a PipecatTool: schema, handler, cancel_on_interruption and register(llm)
voice_processor() a frame processor driving a VoiceAgent
INTEGRATION get_integration("pipecat"); the call id is FunctionCallParams.tool_call_id

Tools on an LLM service

from pipecat.processors.aggregators.llm_context import LLMContext

from agent_adapter.integrations.pipecat import pipecat_tools

tools = pipecat_tools(llm, registry)
context = LLMContext(messages, tools=tools)
When What happens
A tool with effects="additive" or "destructive" registered with cancel_on_interruption=False: an interruption never stops it half-way
A read_only tool or one with no effects cancelled on interruption, Pipecat's default
The model asks for the same booking again (a new call id) the first result returns; one run per LLM service, 120 s window unless your executor sets dedupe_window_s
A tool calls progress("One moment") pushed as a TTSSpeakFrame that stays out of the LLM context
A tool error, denial, guardrail or run limit the result is {"error": {"kind", "message"}}

A VoiceAgent in the pipeline

from pipecat.pipeline.pipeline import Pipeline

from agent_adapter.integrations.pipecat import voice_processor
from agent_adapter.voice import VoiceAgent

processor = voice_processor()
processor.use(VoiceAgent(model, executor, processor.speaker))
pipeline = Pipeline([transport.input(), stt, processor, tts, transport.output()])

UserStartedSpeakingFrame starts a cut and each TranscriptionFrame is a turn or an approval answer; transcripts stop at the processor. Lines go out as TTSSpeakFrames and a cut broadcasts an interruption. Pipecat does not report playout per line, so after a cut the history keeps the lines sent to TTS, which can be a little more than the caller heard.