device-connect-agent-tools

Framework-agnostic tools for Device Connect — discover and invoke devices from any AI agent. Plain Python functions at the core, with adapters for Strands, LangChain, Claude Agent SDK, and MCP.

Contents

• device-connect-agent-tools
  • Contents
  • Where This Fits
  • Install
  • Examples
  • Architecture
    • Hierarchical Discovery
  • Quick Start
    • Plain Python (no framework)
    • Strands Agent
    • LangChain / LangGraph
    • Claude Agent SDK
    • MCP Bridge
  • Connection
    • Auto-Discovery
    • JWT Credentials
    • Explicit Configuration
    • Environment Variables
    • Device-to-Device Mode (No Infrastructure)
  • Tools
  • Event Subscription
  • DeviceConnectMCP — Build Devices with Decorators
  • Writing an Adapter
  • API Reference
    • Connection
    • Tools
    • Connection Object
  • Contributing

Where This Fits

  device-connect-edge          device-connect-server           device-connect-agent-tools
  (Device Connect SDK)    (server runtime)          (agent SDK — this)
        │                         │                         │
        └──────────── Device Connect Mesh ─────────────────────────┘

• device-connect-edge — runs on physical devices (Raspberry Pi, robots, cameras, sensors)
• device-connect-server — runs on servers. Adds registry, security, state, and CLIs
• device-connect-agent-tools — connects AI agents (Strands, LangChain, MCP) to the device mesh

Install

python3 -m venv .venv
source .venv/bin/activate
pip install device-connect-agent-tools

Optional extras:

# Example: install with Strands adapter
pip install -e ".[strands]"

Examples

• StrandsDeviceConnectAgent (device_connect_agent_tools.adapters.strands_agent) — Event-driven Strands Agent that monitors device events and reacts

Architecture

                    ┌──────────────────────────┐
                    │      Your AI Agent       │
                    │  (Strands / LangChain /  │
                    │         MCP)             │
                    └────────┬─────────────────┘
                             │  imports adapter
                    ┌────────▼─────────────────────────┐
                    │  device-connect-agent-tools      │
                    │                                  │
                    │  1. describe_fleet()             │
                    │  2. list_devices(type, location) │  JSON-RPC over
                    │  3. get_device_functions(id)     │  Zenoh / NATS
                    │  4. invoke_device(id, fn, args)  │──────────┐
                    │                                  │          │
                    │  connect() / disconnect()        │          │
                    └──────────────────────────────────┘          │
                                                        ┌─────────▼─────────────┐
                                                        │  Device Connect Mesh  │
                                                        │  ┌─────────────┐      │
                                                        │  │ Camera      │      │
                                                        │  │ Robot       │      │
                                                        │  │ Sensor      │      │
                                                        │  └─────────────┘      │
                                                        └───────────────────────┘

Hierarchical Discovery

Protocols like MCP typically register every device function as a separate tool. With 50 devices averaging 10 functions each, the LLM sees 500 tool definitions — thousands of tokens of JSON Schema that crowd out space for actual reasoning.

This package exposes 4 meta-tools that let the agent drill down progressively, loading detail only when needed:

describe_fleet()                       ~200 tokens  (counts by type and location)
  └─▸ list_devices(device_type=...)    ~50 tokens/device  (names only, no schemas)
        └─▸ get_device_functions(id)   full schemas for ONE device
              └─▸ invoke_device(...)   call a function

Small-fleet shortcut: When the fleet has 5 or fewer devices, describe_fleet() and list_devices() automatically include full function schemas in the response — the agent can skip straight to invoke_device() in one or two calls. The threshold is configurable via the DEVICE_CONNECT_SMALL_FLEET_THRESHOLD environment variable (set to 0 to always require drill-down).

Example — an agent resolving "check the lobby cameras":

Step 1 — describe_fleet() returns a bird's-eye view (~200 tokens):

{
  "total_devices": 47,
  "total_functions": 183,
  "by_type": {
    "camera":  {"count": 12, "locations": ["lobby", "warehouse", "parking"]},
    "robot":   {"count": 8,  "locations": ["warehouse"]},
    "sensor":  {"count": 27, "locations": ["lobby", "warehouse", "parking", "office"]}
  },
  "by_location": {
    "lobby":     {"count": 9,  "types": ["camera", "sensor"]},
    "warehouse": {"count": 31, "types": ["camera", "robot", "sensor"]}
  }
}

Step 2 — Agent sees 12 cameras, narrows to lobby: list_devices(device_type="camera", location="lobby") — compact roster, no schemas:

{
  "devices": [
    {"device_id": "cam-001", "device_type": "camera", "location": "lobby", "function_count": 3, "function_names": ["capture_image", "pan_tilt", "get_status"]},
    {"device_id": "cam-002", "device_type": "camera", "location": "lobby", "function_count": 3, "function_names": ["capture_image", "pan_tilt", "get_status"]}
  ],
  "total": 2, "offset": 0, "limit": 20, "has_more": false
}

Step 3 — Agent picks cam-001, loads its schemas: get_device_functions("cam-001") — full detail for ONE device:

{
  "device_id": "cam-001", "device_type": "camera", "location": "lobby",
  "functions": [{
    "name": "capture_image",
    "description": "Capture a still image",
    "parameters": {"type": "object", "properties": {"resolution": {"type": "string", "enum": ["720p", "1080p", "4k"]}}, "required": ["resolution"]}
  }]
}

Step 4 — invoke_device("cam-001", "capture_image", {"resolution": "1080p"})

Total context: ~450 tokens across 4 calls, vs ~5 000+ if every function schema were loaded upfront. Each step gives the LLM a focused decision ("which type?" → "which device?" → "which function?" → "what params?") instead of a 500-way tool selection.

The same 4 tools are available in every adapter (Strands, LangChain, MCP) and in plain Python. The MCP bridge registers exactly these 4 tools — not one per device function — so MCP clients like Claude Desktop stay responsive regardless of fleet size.

Quick Start

Plain Python (no framework)

from device_connect_agent_tools import (
    connect, disconnect, describe_fleet, list_devices, get_device_functions, invoke_device,
)

connect()

# 1. What's on the network?
fleet = describe_fleet()
print(fleet)  # {total_devices: 47, by_type: {camera: {count: 12}, ...}}

# 2. Browse a specific type
cameras = list_devices(device_type="camera", location="lobby")
for d in cameras["devices"]:
    print(f"{d['device_id']}: {d['function_names']}")

# 3. Get full schemas for one device
info = get_device_functions("cam-001")

# 4. Call a function
result = invoke_device("cam-001", "capture_image", {"resolution": "1080p"})
print(result)

disconnect()

Strands Agent

from strands import Agent
from strands.models import AnthropicModel
from device_connect_agent_tools import connect
from device_connect_agent_tools.adapters.strands import (
    describe_fleet, list_devices, get_device_functions, invoke_device,
)

connect()

agent = Agent(
    model=AnthropicModel(model_id="claude-sonnet-4-20250514"),
    tools=[describe_fleet, list_devices, get_device_functions, invoke_device],
    system_prompt="You manage devices on a Device Connect network.",
)

agent("Find all cameras in the lobby and capture an image from each one.")

LangChain / LangGraph

from langchain_anthropic import ChatAnthropic
from langgraph.prebuilt import create_react_agent
from device_connect_agent_tools import connect
from device_connect_agent_tools.adapters.langchain import (
    describe_fleet, list_devices, get_device_functions, invoke_device,
)

connect()

model = ChatAnthropic(model="claude-sonnet-4-20250514")
agent = create_react_agent(model, tools=[describe_fleet, list_devices, get_device_functions, invoke_device])

result = agent.invoke({"messages": [{"role": "user", "content": "What devices are online?"}]})
print(result["messages"][-1].content)

Claude Agent SDK

import anyio
from claude_agent_sdk import ClaudeSDKClient, ClaudeAgentOptions, AssistantMessage, TextBlock
from device_connect_agent_tools import connect
from device_connect_agent_tools.adapters.claude import create_device_connect_server

async def main():
    connect()
    options = ClaudeAgentOptions(
        model="claude-sonnet-4-6",
        mcp_servers={"device_connect": create_device_connect_server()},
    )
    async with ClaudeSDKClient(options=options) as client:
        await client.query("What devices are online?")
        async for message in client.receive_response():
            if isinstance(message, AssistantMessage):
                for block in message.content:
                    if isinstance(block, TextBlock):
                        print(block.text)

anyio.run(main)

MCP Bridge

The MCP bridge exposes the same 4 hierarchical meta-tools over the MCP protocol. Unlike traditional MCP servers that register one tool per device function, the bridge stays constant regardless of fleet size.

Add to your MCP client config (e.g., ~/Library/Application Support/Claude/claude_desktop_config.json):

{
    "mcpServers": {
        "device-connect": {
            "command": "python",
            "args": ["-m", "device_connect_agent_tools.mcp"],
            "env": {
                "ZENOH_CONNECT": "tcp/localhost:7447",
                "DEVICE_CONNECT_ALLOW_INSECURE": "true"
            }
        }
    }
}

The MCP client sees describe_fleet, list_devices, get_device_functions, and invoke_device — the agent navigates the fleet the same way it would via Strands or LangChain.

Connection

Auto-Discovery

connect() with no arguments works when run from anywhere inside a Device Connect project tree:

from device_connect_agent_tools import connect

connect()  # auto-discovers broker URL, credentials, and TLS certs

It searches upward from the current directory for security_infra/credentials/ and security_infra/certs/.

JWT Credentials

Agents need their own credentials generated by device-connect-server:

# On the server: generate agent credentials
./security_infra/gen_creds.sh --user my-agent

Then connect using the credentials file:

MESSAGING_CREDENTIALS_FILE=~/.device-connect/credentials/my-agent.creds.json \
  NATS_URL=nats://localhost:4222 python my_agent.py

Or pass credentials explicitly:

connect(
    messaging_urls=["nats://localhost:4222"],
    credentials={"jwt": "...", "nkey_seed": "..."},
)

Explicit Configuration

connect(
    messaging_urls=["tls://nats.example.com:4222"],
    zone="production",
    credentials={"jwt": "...", "nkey_seed": "..."},
    tls_config={"ca_file": "/path/to/ca.pem"},
)

Environment Variables

Resolution order: explicit parameter > environment variable > auto-discovery.

Device-to-Device Mode (No Infrastructure)

With no endpoint URLs configured, the discovery tools automatically use D2D presence-based discovery (Zenoh multicast scouting) instead of querying the registry service. No Docker infrastructure needed:

export DEVICE_CONNECT_ALLOW_INSECURE=true
# No ZENOH_CONNECT → Zenoh multicast scouting on LAN (default)
python my_agent.py

Devices on the same LAN that are also running in D2D mode will be discovered automatically. See device-connect-edge for details.

Tools

The 4 hierarchical discovery tools are plain Python functions. Import them directly or from an adapter:

# Plain Python
from device_connect_agent_tools import describe_fleet, list_devices, get_device_functions, invoke_device

# Strands (returns DecoratedFunctionTool)
from device_connect_agent_tools.adapters.strands import describe_fleet, list_devices, get_device_functions, invoke_device

# LangChain (returns StructuredTool)
from device_connect_agent_tools.adapters.langchain import describe_fleet, list_devices, get_device_functions, invoke_device

Event Subscription

For event-driven agents that react to device events in real-time:

import asyncio
from device_connect_agent_tools import connect, get_connection

connect()
conn = get_connection()

async def on_event(msg):
    print(f"Event on {msg.subject}: {msg.data.decode()}")

# Subscribe to all device events
sub = await conn.async_subscribe("device-connect.default.*.event.>", on_event)

Subject patterns:

• device-connect.{zone}.*.event.> — all events from all devices
• device-connect.{zone}.{device_id}.event.> — events from one device
• device-connect.{zone}.{device_id}.event.words_generated — specific event type

DeviceConnectMCP — Build Devices with Decorators

DeviceConnectMCP provides a FastMCP-compatible API for building Device Connect devices with @tool and @event decorators. Auto-generates function schemas from type hints and docstrings.

from device_connect_agent_tools.mcp import DeviceConnectMCP

mcp = DeviceConnectMCP(
    "cleaning-robot-001",
    device_type="cleaning_robot",
    manufacturer="Acme",
    location="warehouse-A",
)

@mcp.tool()
async def start_cleaning(zone: str = "all") -> dict:
    """Start cleaning in the specified zone."""
    return {"status": "started", "zone": zone}

@mcp.event()
async def cleaning_complete(zone: str, duration_seconds: int):
    """Emitted when cleaning is complete."""
    pass

await mcp.run()

Writing an Adapter

To add support for another framework, wrap the plain functions:

# device_connect_agent_tools/adapters/my_framework.py

from my_framework import wrap_tool
from device_connect_agent_tools.tools import (
    describe_fleet as _describe_fleet,
    list_devices as _list_devices,
    get_device_functions as _get_device_functions,
    invoke_device as _invoke_device,
)

describe_fleet = wrap_tool(_describe_fleet)
list_devices = wrap_tool(_list_devices)
get_device_functions = wrap_tool(_get_device_functions)
invoke_device = wrap_tool(_invoke_device)

API Reference

Connection

Hierarchical Discovery Tools

Additional Tools

Connection Object

For advanced use via get_connection():

Contributing

We welcome contributions! Please open an issue to report bugs or suggest features, or submit a pull request directly.