livelink 0.1.1

Supervised runtime for voice AI agents — Gemini Live, OpenAI Realtime, one interface

LiveLink

The supervised runtime for voice AI agents.

5 lines to a working voice agent. Add tools in +3. Add guardrails in +5. Grow to multi-agent without rewriting.

from livelink import LiveAgent

agent = LiveAgent(
    model="gemini/gemini-2.5-flash-native-audio",
    instructions="You are a helpful voice assistant.",
)
agent.serve()  # opens http://localhost:8000 — talk to your agent

Install

pip install livelink[gemini,serve]   # Gemini Live
pip install livelink[openai,serve]   # OpenAI Realtime
pip install livelink[all]            # All providers + serve

Requires Python 3.10+.

Quickstart

Get a Google AI API key, then:

export GOOGLE_API_KEY=your-key-here

from livelink import LiveAgent

agent = LiveAgent(
    model="gemini/gemini-2.5-flash-native-audio",
    instructions="You are a helpful voice assistant.",
)
agent.serve()

Open http://localhost:8000. Talk to your agent.

Add a Tool (+3 lines)

from livelink import LiveAgent

agent = LiveAgent(
    model="gemini/gemini-2.5-flash-native-audio",
    instructions="You help with weather questions.",
)

@agent.tool
async def get_weather(city: str) -> str:
    """Get current weather for a city."""
    return f"72°F and sunny in {city}"

agent.serve()

Tool schemas are inferred automatically from type hints and docstrings.

Add Guardrails (+5 lines)

from livelink import LiveAgent, input_guardrail, GuardrailResult

@input_guardrail
async def block_pii(content: str) -> GuardrailResult:
    if "@" in content or "SSN" in content:
        return GuardrailResult(action="block", message="PII detected")
    return GuardrailResult(action="pass")

agent = LiveAgent(
    model="gemini/gemini-2.5-flash-native-audio",
    instructions="You are a helpful voice assistant.",
    input_guardrails=[block_pii],
)
agent.serve()

Guardrails run at every input/output boundary. Chain multiple, modify content, or block outright.

Multi-Agent Handoffs (+8 lines)

from livelink import LiveAgent, Handoff

billing = LiveAgent(
    model="gemini/gemini-2.5-flash-native-audio",
    instructions="You are a billing specialist.",
)

agent = LiveAgent(
    model="gemini/gemini-2.5-flash-native-audio",
    instructions="You are a support router. Transfer billing questions.",
    handoffs=[Handoff(target=billing, tool_name="transfer_to_billing")],
)
agent.serve()

The transport stays open. Conversation history accumulates across swaps.

What's Included

Providers

• Gemini Live (Google GenAI SDK)
• OpenAI Realtime API
• One interface for both — swap model= to switch

Runtime

• agent.serve() — zero-config WebSocket server + browser UI
• Runner.run(agent, transport) — structured lifecycle with callbacks
• agent.session().run(transport) — full manual control
• All three use the same underlying session — no rewrite cliffs

Supervision

• Observable tool calls, audio events, and reasoning steps via EventBus
• Human-in-the-loop (HITL) input requests with priority, timeout, cancellation
• Background task management with cancellation support
• Conversation state machine (idle → active → interrupted → complete)

Agent patterns

• @agent.tool — automatic JSON schema from type hints + docstrings
• @input_guardrail / @output_guardrail — content validation at boundaries
• Handoff — control transfer between agents (transport stays open)
• agent.as_tool() — delegation pattern (sub-agent runs as a tool call)

LangGraph integration

pip install livelink[langchain]

Run any LangGraph graph as a voice agent with full interrupt/resume support.

Architecture

One runtime path, progressively revealed:

agent.serve()                          # 5 lines — zero config
Runner.run(agent, transport, on_*=…)   # 15 lines — callbacks + lifecycle
agent.session(config=…).run(transport) # 30+ lines — full control

LiveAgent is pure config. RealtimeSession is the runtime. Runner manages the lifecycle between them. Adapters translate provider-specific protocols — your code never sees the difference.

Examples

See examples/ for runnable demos:

• minimal_voice/ — simplest possible agent
• pizza_order/ — tool use with structured state
• supervised_agent/ — HITL approvals and supervisor injection
• escalation_handler/ — multi-step investigation with graduated autonomy
• multi_agent/ — handoffs and delegation
• guardrails/ — input/output content filtering
• incident_response/ — LangGraph ReAct agent with voice interface

License

MIT