pinecall

Build voice and chat agents as Python classes. Annotated fields are the agent's state, @tool methods are what the model can call, docstrings are what it reads about them, and a Jinja view renders the part of the prompt that changes during the call.

from typing import Literal

from pinecall import Agent, state, tool

from clinica import agenda  # your own system: a CRM, an API, a database


class ClinicaNorte(Agent):
    """You are the front desk of Clínica Norte. Formal, short sentences."""

    stage: Literal["identify", "book"] = "identify"
    patient: dict[str, str] | None = state(None, pii=True)
    slots: list[dict[str, str]] = []

    @tool(stage="identify", pii=("name", "phone"))
    def find_patient(self, name: str, phone: str) -> dict[str, str] | None:
        """Finds the patient by name and phone. Ask for both before calling it."""
        self.patient = agenda.find(name, phone)
        if self.patient:
            self.stage = "book"
        return self.patient

    @tool(stage="book", preview=2)
    def free_slots(self, day: str) -> list[dict[str, str]]:
        """Free slots on one day."""
        self.slots = agenda.free(day)
        return self.slots

views/clinica-norte.jinja, beside the class, renders the prompt from that state:

{% if stage == "identify" %}
Greet the caller and ask for their name and phone. Nothing else until the patient is identified.
{% endif %}
{% if remembers("usual doctor") %}
Offer their usual doctor's slots first.
{% endif %}
{% if slots %}
## Free slots, in order

{% for slot in slots %}
{{ slot.when }} with {{ slot.doctor }}
{% endfor %}
{% endif %}

The prompt has four named blocks: identity, knowledge and tools are static and cached before the history; view comes after it and is replaced every turn. A block is re-sent only when its text changes.

Memory and knowledge-base results never enter the prompt. The platform runs recall and search and returns their results to the model as tool results, so retrieved text is treated as data, not instructions. A view can ask remembers("usual doctor") and write its own sentence about the answer.

Five minutes

uv add pinecall, or pip install pinecall. PyPI pinecall 0.1.0 to 0.1.5 are the runtime's releases, before it moved to pinecall-runtime: a project pins pinecall>=0.1.6.

This package is a library: the verbs are the one pinecall CLI's, the same for every language. In a project laid out as agents/<name>/agent.py, with this package in its .venv:

npm i -g pinecall                     # the CLI: Node, whatever the agent is written in
pinecall login                        # this machine signed in, once; pinecall.json names the org
pinecall prompt --state agents/<name>/test/goldens/a.json   # the exact prompt a state produces. No gateway.
pinecall start                        # registered and answering: the process you deploy; leave it running
pinecall chat                         # in another terminal: a caller against the agent start holds
pinecall test                         # ring 1: the goldens, played on the agent start holds, scored by the gateway

The CLI never loads your class: for prompt and start it runs this package's serve entry, .venv/bin/python -m pinecall.serve, and talks to it through the gateway. start is the one verb that runs the agent; chat, test and simulate start nothing, and reach whoever holds it. The console, the knowledge bases, memory and every other verb are the CLI's. How a server runs the agent, with the CLI or inside your own Python process, is docs/production.md.

Architecture

This package is the application side. It never talks to a model vendor or handles audio: it sends commands and reads log entries over one WebSocket, in the runtime's own wire (pinecall.wire), held to the runtime's golden call log. Tools and hooks are the only code that may change state.

import asyncio
import os

from pinecall import Client
from pinecall.wire.events import UserTurnEnded


async def main() -> None:
    pc = Client("https://cloud.pinecall.io", os.environ["PINECALL_KEY"])
    agent = pc.agent("clinica-norte")

    def heard(event, call):  # every turn.user of every call this socket serves
        if isinstance(event, UserTurnEnded):
            call.say(f"I heard: {event.text}")

    agent.on("turn.user", heard)
    await pc.connect()
    await asyncio.Event().wait()


asyncio.run(main())

PINECALL_KEY here is a server's token, made in the console's API keys screen for the org and the environment the process runs in — not your own sign-in, which stays on your machine. On a runtime of your own the URL is its gateway's (http://127.0.0.1:8080 on your machine).

Testing it

Agent tests run with no network, key, model or gateway, in a plain pytest:

from pinecall.testing import Gateway, load

ClinicaNorte = load("agents/clinica-norte/agent.py")

with Gateway() as pc:
    pc.mount(ClinicaNorte)
    call = pc.call_started(from_="+34600123456")
    call.tool("find_patient", name="Marta Ruiz", phone="600123456")

    assert "ask for their name" not in call.prompt
    assert call.tools == ["free_slots"]

Where the rest is

Developing

uv sync
make check        # ruff, pyright strict, deptry, the suite and the example's own

Requirements

Python 3.11+. Runtime dependencies: pydantic, Jinja2, websockets, httpx and typing-extensions. The pinecall CLI is a Node program (Node 24+): the verbs need it, the library does not.

License

Apache-2.0.