voicelayer-agents

VoiceLayer channels for the Python agent framework you already use: SMS, WhatsApp, Telegram, Discord and email, for the Claude Agent SDK, the OpenAI Agents SDK and LangChain/LangGraph.

• Tool packs — VoiceLayer's channel tools (message_send, message_reply, conversations_list, conversation_read, conversation_set_controller, messages_wait, channels_list, whatsapp_templates_list) in each framework's own MCP hook. The VoiceLayer MCP server lists them; nothing is re-implemented here.
• serve_direct — direct mode, with any framework: your process answers the conversations handed to code (controller api). Your runner gets the message and the conversation's history and returns the reply; it is sent once, through every gate of the pipeline (the Output Guard included — a refusal goes to on_refused, never retried around).

pip install voicelayer-agents              # serve_direct and the test-mode helper
pip install "voicelayer-agents[openai]"    # + the OpenAI Agents tool pack; also [claude], [langgraph]

Each framework is an extra, pinned to the version the tests run. Direct mode is built on the generated API client, voicelayer, installed with it.

Keys

Set VOICELAYER_API_KEY to a workspace API key (dashboard -> API keys) scoped to what the agent does: messages:send, messages:read, conversations:control, events:read, channels:read. Every helper reads it, and VOICELAYER_API_URL (default https://api.vlayers.ai), so the examples below pass nothing. api_key= and base_url= override both. A test key (vl_test_...) runs everything in test mode: every send goes through the loopback and costs nothing, and a workspace with no number gets a free test sender. A worker key (vlw_...) is refused.

Tool packs

import asyncio

from agents import Agent, Runner
from voicelayer_agents.openai import voicelayer_mcp_server


async def main():
    async with voicelayer_mcp_server() as voicelayer:  # connected; closed afterwards
        agent = Agent(name="support", instructions="You answer customers in one short sentence.", mcp_servers=[voicelayer])
        print((await Runner.run(agent, 'Text "Hello from VoiceLayer" to +15550101 over sms.')).final_output)


asyncio.run(main())

from claude_agent_sdk import ClaudeAgentOptions
from voicelayer_agents.claude import voicelayer_mcp

options = ClaudeAgentOptions(**voicelayer_mcp(), tools=[], system_prompt="You answer customers in one short sentence.")

from langgraph.prebuilt import create_react_agent
from voicelayer_agents.langgraph import voicelayer_tools

agent = create_react_agent("openai:gpt-4.1-mini", await voicelayer_tools())  # inside an async function

Direct mode

from voicelayer_agents import DirectAnswer, serve_direct


def answer(turn):
    # turn.message: the inbound (id, channel, conversation_id, text); turn.history: the conversation, oldest first
    if "person" in turn.message.text.lower():
        return DirectAnswer(text="Passing you to a colleague.", handoff="asked for a person")
    return f"You said: {turn.message.text}"


serve_direct(answer).wait()  # answers until Ctrl+C

serve_direct returns once its place in the event stream is fixed: every message that arrives after it returns is answered. .wait() blocks (a program that only calls serve_direct would exit at once); .stop() or a with block ends it. Pass after= to resume from a saved cursor, channels= to narrow, history=False (or your own callable) to skip the history read.

Test mode

from voicelayer_agents import simulate_inbound

# a person writing to the workspace, simulated (test keys only): sms, from +15550101 unless you pass channel= / from_=
simulate_inbound("Are you open?")

Guides: vlayers.ai/docs/agents. The TypeScript twin is @voicelayer/agents. Licence: Apache-2.0.