How the Port OS MCP integration optimizes UX for Agents
Published October 20, 2026 · By Port OS Team · Drafted with AI
MCP, the Model Context Protocol, is a standard way for an AI client to call tools that a remote server offers. Port OS runs an MCP server, so any client that supports remote MCP servers with OAuth sign-in can work with your devices.
This post gives the overview. The MCP guide has the setup steps and the MCP reference has the full tool list, and we do not repeat them here.
Step 1: the client connects
You add the Port OS server URL in your AI client's connector settings. The client then signs in on your behalf using OAuth. Clients that support dynamic client registration or client ID metadata documents can register themselves. For clients that support neither, you can register an OAuth client yourself.
Step 2: you grant access
During sign-in you review the permissions and choose what the assistant gets. Reading your device list and controlling devices are separate permissions. The assistant gets what you approve and nothing more. Port OS checks access rules on each operation, not only at sign-in.
The guide How to authorize access with scopes and resource permissions explains the model.
Step 3: the client discovers its tools
The client asks the server which tools it offers. Port OS offers tools to find devices, send commands, manage your device list and groups, and look at event subscriptions. The exact set can vary by deployment, so the client reads it from the server instead of assuming it.
The main two are:
- A query tool that lists your devices and shows what each supports, including the exact message IDs it accepts.
- A control tool that sends one of those messages to a device or group, with a payload that follows the message's schema.
The MCP tool reference lists all of them.
Step 4: the assistant acts and checks the result
When you say "turn off the simulator light", the assistant queries your devices, finds the light and copies the message ID for turning it off from the results. It then sends that message through the control tool.
Capabilities and messages come from the device's declared interface, so the assistant is not guessing at a vendor's commands. You can browse the same definitions in the capability reference.
An accepted command is not proof that a physical device changed. Port OS guides recommend that an assistant reads fresh state afterwards, and treats a stale or unavailable reading as unknown rather than as "off" or "offline".
Guides for agents
Port OS also publishes its task guides as Markdown that an agent can read, in the agent guide library. Clients that only support tools can use a bundled skill that carries the same guides. The guides are documentation resources and not extra tools.
Events
Event subscriptions let an assistant react when something happens, such as a delivery arriving at your door. They work only in ChatGPT today, and only for accounts on the waitlist. How to react to device events explains how they work.
Where to go next
- Quick start: add a simulator light and try it.
- Connect your AI client: guides for ChatGPT, Claude, Muse, Grok, Codex and custom clients.
- Agentic API development: top 10 best practices: the practices behind this design.