connexity-pipecat 0.1.1

AI outbound voice agent framework

Connexity Pipecat

Connexity Pipecat is a flexible voice‑AI agent framework that pairs the audio‑first power of Pipecat with Twilio telephony and an LLM of your choice (OpenAI, Gemini, Groq, Fireworks, …). Use it to spin up production‑grade phone agents that can talk, listen and book meetings—all in real time.

---

Contents

• Installation
  • From PyPI
  • Local development
• Repository layout
• Quick start
• Creating your own agent
  • Project scaffold
  • Configuration file (config.yaml)
    • Dynamic variables (generators)
    • Registering custom tools
    • Registering custom handlers
  • Built‑in handlers and processors
    • Generators
    • Tools
    • Handlers
• Environment variables
• License

---

Installation

From PyPI

pip install connexity-pipecat

Local development

git clone /link/
cd connexity-pipecat
pip install -e ".[dev,docs]"

---

Repository layout

src/connexity_pipecat/
├── core/                 # Runtime core: agents, config, LLMs, tools
│   ├── tools/            # Built‑in tool registry + helpers
│   ├── generators/       # Data generators for dynamic prompt values
│   ├── voice_calls/      # Twilio integration helpers & templates
│   ├── config.py         # YAML loader + caching
│   └── ...
├── api/                  # FastAPI routers & handler glue
├── data/                 # Pydantic schemas, cache, constants
└── assets/               # Background audio & SFX

---

Quick start

from pathlib import Path

from dotenv import load_dotenv
from fastapi import FastAPI
from fastapi.middleware.cors import CORSMiddleware

from connexity_pipecat import init_config, register_routes

load_dotenv()

PROJECT_ROOT = Path(__file__).parent
init_config(str(PROJECT_ROOT / "config.yaml"))

app = FastAPI()
register_routes(app)

app.add_middleware(
    CORSMiddleware,
    allow_origins=["*"],
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

Start the server:

uvicorn main:app --reload

---

Creating your own agent

Project scaffold

my-agent/
├── main.py
├── .env
├── config.yaml
└── custom/
    ├── handlers/
    │   └── say_hi.py
    └── tools/
        ├── config.yaml
        └── functions/
            └── greet_user.py

Configuration file (config.yaml)

Everything is driven by a single YAML file. A trimmed example:

llm:
  main: {vendor: openai, model: gpt-4o-mini}

agent_inputs:
  agent_name: Emma
  current_timestamp: {generator: current_timestamp}  # evaluated at runtime

tools: ["end_call"]

routes:
  routers:
    - prefix: ""
      routes:
        - {path: /, methods: [POST], handler: platform_inference}
  websockets:
    - prefix: ""
      routes:
        - {path: /outbound/ws, handler: outbound_websocket_endpoint}

Full key reference

Key (dot-notation)	Type	Allowed / typical values	Notes
vad_params.confidence	float	0.0 – 1.0	Voice-activity detector (higher = stricter).
vad_params.min_volume	float	0.0 – 1.0	Minimum RMS loudness considered speech.
vad_params.start_secs	float	> 0.0	Silence threshold (s) before speech start.
vad_params.stop_secs	float	> 0.0	Silence threshold (s) before speech end.
llm.main.vendor	str	openai, groq, fireworks, google, …	LLM backend for primary inference.
llm.main.model	str	Model name	e.g. gpt-4o-mini, mixtral-8x7b.
llm.utils.vendor	str	same as above	Lightweight helper LLM.
llm.utils.model	str	Model name	Used for summaries, embeddings, etc.
use_connexity_observer	bool	true / false	Push call metrics to Connexity API.
pipeline_settings.audio_in_sample_rate	int	8000 – 48000	Incoming PCM sample-rate (Hz).
pipeline_settings.allow_interruptions	bool	true / false	Let user barge-in over TTS.
pipeline_settings.enable_metrics	bool	true / false	Collect latency & token stats.
pipeline_settings.report_only_initial_ttfb	bool	true / false	Emit only first-token metrics.
vector_db.type	str	Weaviate, Chroma, None	External store for long-term memory.
embedding.type	str	Provider/model id	e.g. OpenAI/text-embedding-3-small.
agent_inputs.project_name	str	lowercase slug	Logical project grouping.
agent_inputs.language_code	ISO 639-1	en, de, es, …	Used for STT/LLM model selection.
agent_inputs.language	str	Language name	Display-only.
agent_inputs.translate_prompt	bool	true / false	Auto-translate system prompt.
agent_inputs.agent_name	str	Human-friendly	“Emma”, “Max”, etc.
agent_inputs.agent_company_name	str	Free text	Company brand in prompts.
agent_inputs.*.generator	str	Function name	Any registered generator (see below).
tools	list[str]	Built-ins + custom	e.g. ["end_call", "book_appointment"].
agent_id	str	slug/UUID	Internal analytics identifier.
routes.routers[].prefix	str	Path prefix	Empty string = root.
routes.routers[].routes[].path	str	URL path	Must match FastAPI route.
routes.routers[].routes[].methods	list[str]	["GET"], ["POST"], …	HTTP verbs.
routes.routers[].routes[].handler	str	Handler name	From connexity_pipecat.api.handlers or custom.
routes.websockets[].prefix	str	Path prefix	Usually empty.
routes.websockets[].routes[].path	str	URL path	WebSocket endpoint.
routes.websockets[].routes[].handler	str	Handler name	From api.handlers or custom.
start_message	str	Short greeting	Spoken when call connects (optional).
prompts.agent.<lang>	str (multiline)	Any	Primary system prompt.
prompts.quick_response.<lang>	str (multiline)	Any	Prompt for filler-phrase generator.
prompts.post_analysis.<lang>	str (multiline)	Any	Prompt for call summariser.

^{Dot-notation shows nested keys; array items are marked with [].}

Dynamic variables (generators)

Any scalar can be generated dynamically with a function. Built‑ins include:

Generator	What it returns
current_timestamp	"YYYY‑MM‑DD, Monday. Time: 15:04."
get_available_time_slots_str	JSON list of free slots for next business days

Add your own generator and register it with register_custom_generators().

Registering custom tools

1. Describe your tool in custom/tools/config.yaml.
2. Implement it in custom/tools/functions/.
3. Register them:

register_custom_tools("custom/tools/functions", "custom/tools/config.yaml")

Registering custom handlers

Drop additional FastAPI handlers in custom/handlers/ and register:

register_custom_handlers("custom/handlers")

Built‑in handlers and processors

Generators (for computed agent_inputs)

• current_timestamp
• get_available_time_slots_str
• get_grouped_calendar

Tools

• transfer_call
• end_call
• get_weekday
• get_available_time_slots
• await_call_transfer
• book_appointment

Handlers

connexity_pipecat.api.handlers exposes reusable FastAPI coroutine handlers for both HTTP and WebSocket endpoints.
You may add them to the routes section of your config.yaml as shown in the sample project.

Path	Method(s)	Handler name	Purpose
/	POST	platform_inference	WNH platform text-only inference
/initiate_phone_call	POST	initiate_phone_call	Kick off an outbound Twilio call
/voice_updates	POST	receive_voice_updates	Update response-ID & turn-taking flags from the client
/call_status/{sid}	GET	status_callback_get	Poll current call status
/call_status	POST	status_callback_post	Twilio StatusCallback webhook
/outbound/webhook	POST	outbound_webhook	Return TwiML to connect Twilio ↔ WebSocket (outbound)
/inbound/webhook	POST	inbound_webhook	Return TwiML to connect Twilio ↔ WebSocket (inbound)

WebSocket endpoints

Path	Handler name	Purpose
/outbound/ws	outbound_websocket_endpoint	Media stream for outbound calls
/inbound/ws	inbound_websocket_endpoint	Media stream for inbound calls

Add the routes to your config.yaml like so:

routes:
  routers:
    - prefix: ""
      routes:
        - {path: /, methods: [POST], handler: platform_inference}
        - {path: /initiate_phone_call, methods: [POST], handler: initiate_phone_call}
        - {path: /voice_updates, methods: [POST], handler: receive_voice_updates}
        - {path: /call_status/{sid}, methods: [GET], handler: status_callback_get}
        - {path: /call_status, methods: [POST], handler: status_callback_post}
        - {path: /outbound/webhook, methods: [POST], handler: outbound_webhook}
        - {path: /inbound/webhook, methods: [POST], handler: inbound_webhook}

  websockets:
    - prefix: ""
      routes:
        - {path: /outbound/ws, handler: outbound_websocket_endpoint}
        - {path: /inbound/ws, handler: inbound_websocket_endpoint}

---

Environment variables

Create a .env next to main.py:

###############################################
# Connexity-Pipecat – Environment Variables   #
# Copy this file to `.env` and fill the blanks #
###############################################

##########################
# LLM provider API keys  #
##########################
OPENAI_API_KEY=
GROQ_API_KEY=
FIREWORKS_API_KEY=
# Google-Vertex / Gemini
GOOGLE_CREDENTIALS_JSON=

##########################
# Voice & speech         #
##########################
DEEPGRAM_API_KEY=
ELEVENLABS_API_KEY=
ELEVENLABS_VOICE_ID=

##########################
# Calendar / tools hooks #
##########################
TOOL_CHECK_SLOT_AVAILABILITY_WEBHOOK_URL=
TOOL_BOOKING_WEBHOOK_URL=

##########################
# Twilio telephony       #
##########################
TWILIO_ACCOUNT_ID=
TWILIO_AUTH_TOKEN=
TWILIO_PHONE_NUMBER=[phone]

####################################
# Public hostname of your FastAPI  #
####################################
# e.g. "voice.example.com" (no http://, no trailing slash)
SERVER_ADDRESS=

###########################################
# Optional call-status callback endpoint  #
###########################################
CALL_STATUS_URL=

###########################################################
# HTTP inference micro-service (text chat) & Connexity API #
###########################################################
INFERENCE_URL=
CONNEXITY_API_KEY=

---

License

MIT – see LICENSE.