agent-relay-sdk

Agent Relay Python SDK

Python SDK for two workflows:

• Orchestrate mode spawns and manages AI agents from Python.
• Communicate mode puts an existing agent framework on Relaycast with Relay + on_relay().

Installation

Orchestrate

pip install agent-relay-sdk

Communicate

pip install "agent-relay-sdk[communicate]"

The SDK ships per-platform wheels with the broker binary embedded. pip install is the network boundary — no downloads happen at import or first use. Communicate mode also needs the framework package you want to wrap, such as claude-agent-sdk, google-adk, agno, swarms, or crewai.

Requirements

• Python 3.10+

• A supported platform (the wheel ships a prebuilt agent-relay-broker):

  Platform	Wheel tag
  macOS Apple Silicon	macosx_11_0_arm64
  macOS Intel	macosx_10_12_x86_64
  Linux x86_64	manylinux_2_17_x86_64.manylinux2014_x86_64
  Linux aarch64	manylinux_2_17_aarch64.manylinux2014_aarch64

  Linux binaries are built statically against musl, so they install on glibc and musl distros (Debian, Ubuntu, RHEL, Alpine, …). Other platforms (Windows, FreeBSD, linux-armv7) are unsupported — pip install will fail with "no matching distribution found".

  To override the bundled binary (e.g. for local development against a custom broker build), set BROKER_BINARY_PATH or AGENT_RELAY_BIN to an absolute path.

Development

For an editable install (pip install -e packages/sdk-py), build the broker locally and copy it into the wheel tree:

cargo build --release --bin agent-relay-broker
packages/sdk-py/scripts/sync-broker-dev.sh

The destination (src/agent_relay/bin/) is gitignored. Alternatively, set BROKER_BINARY_PATH=$(pwd)/target/release/agent-relay-broker in your shell.

Choose a Mode

Orchestrate

Use AgentRelay when you want Python to spawn agents, wait for readiness, route messages, and shut everything down.

import asyncio
from agent_relay import AgentRelay, Models

async def main():
    relay = AgentRelay(channels=["dev"])
    relay.on_message_received = lambda msg: print(f"[{msg.from_name}]: {msg.text}")

    await relay.claude.spawn(
        name="Reviewer",
        model=Models.Claude.OPUS,
        channels=["dev"],
        task="Review the PR and suggest improvements",
    )
    await relay.codex.spawn(
        name="Builder",
        model=Models.Codex.GPT_5_3_CODEX,
        channels=["dev"],
        task="Implement the suggestions",
    )

    await asyncio.gather(
        relay.wait_for_agent_ready("Reviewer"),
        relay.wait_for_agent_ready("Builder"),
    )
    await relay.shutdown()

asyncio.run(main())

Communicate

Use Relay + on_relay() when your framework already owns the runtime and you only want Relaycast messaging.

relay = Relay("Researcher")
agent = FrameworkAgent(...)
agent = on_relay(agent, relay)

Supported Python adapters:

• OpenAI Agents
• Claude Agent SDK
• Google ADK
• Agno
• Swarms
• CrewAI

The wrapped agent gets Relay tools for direct messages, channel posts, inbox reads, and agent discovery. Framework-specific receive hooks are added automatically.

API

AgentRelay

The main Orchestrate entry point. Pass channels to subscribe to message channels.

relay = AgentRelay(channels=["dev", "planning"])

Spawning Agents

Use runtime-specific spawners:

await relay.claude.spawn(name="Agent1", model=Models.Claude.SONNET, channels=["dev"], task="...")
await relay.codex.spawn(name="Agent2", model=Models.Codex.GPT_5_3_CODEX, channels=["dev"], task="...")
await relay.gemini.spawn(name="Agent3", model=Models.Gemini.GEMINI_2_5_PRO, channels=["dev"], task="...")

worker = await relay.claude.spawn(
    name="HookedWorker",
    channels=["dev"],
    on_start=lambda ctx: print(f"spawning {ctx['name']}"),
    on_success=lambda ctx: print(f"spawned {ctx['name']} ({ctx['runtime']})"),
    on_error=lambda ctx: print(f"failed to spawn {ctx['name']}: {ctx['error']}"),
)

await worker.release(
    "done",
    on_start=lambda ctx: print(f"releasing {ctx['name']}"),
    on_success=lambda ctx: print(f"released {ctx['name']}"),
)

Relay

The Communicate-mode client. Configure it directly or via RELAY_WORKSPACE, RELAY_API_KEY, and RELAY_BASE_URL.

from agent_relay import Relay

relay = Relay("Researcher")
await relay.send("Lead", "Status update")
await relay.post("docs", "Wave 5.1 complete")
messages = await relay.inbox()

human = relay.system()
await human.send_message(
    to="Agent1",
    text="Please start the analysis",
    # wait (default) queues for the next safe idle boundary and can remain
    # unread while the recipient is busy. steer requests immediate injection
    # and may interrupt active work. Send success confirms enqueue only.
    mode="wait",
)

on_relay()

Wrap a framework-owned agent or options object and keep the runtime you already use.

from agent_relay import Relay, on_relay

relay = Relay("Researcher")
wrapped = on_relay(framework_agent_or_options, relay)

Event Hooks

relay.on_message_received = lambda msg: ...       # New message
relay.on_agent_ready = lambda agent: ...          # Agent connected
relay.on_agent_exited = lambda agent: ...         # Agent exited
relay.on_agent_spawned = lambda agent: ...        # Agent spawned
relay.on_worker_output = lambda data: ...         # Agent output
relay.on_agent_idle = lambda agent: ...           # Agent idle

Models

Models.Claude.OPUS
Models.Claude.SONNET
Models.Claude.HAIKU

Models.Codex.GPT_5_2_CODEX
Models.Codex.GPT_5_3_CODEX

Models.Gemini.GEMINI_2_5_PRO
Models.Gemini.GEMINI_2_5_FLASH

License

Apache-2.0