AI Agent Access for Security Cameras: How MCP Connects Your Camera to Claude, ChatGPT, and More

Oriel is the first security camera app that lets AI agents access your camera feed. Using Model Context Protocol (MCP), agents like Claude, ChatGPT, or any MCP-compatible tool can check camera status, grab snapshots, and review detection events. Everything runs on your local network. No cloud involved.

What This Means in Practice

You're working at your desk. Instead of picking up your phone, opening an app, and checking the camera yourself, you type:

"Check the front door camera."

Your AI agent calls the camera, receives a snapshot, and tells you what it sees. A closed door, a package on the mat, a person at the gate. The agent uses its built-in vision capabilities to interpret the image, then reports back in plain language.

This works with any MCP-compatible AI tool: Claude Code, Claude Desktop, ChatGPT with plugins, or custom agents built on the MCP standard.

How It Works

Oriel's camera device runs a small HTTP server on your local network. An MCP server bridges the gap between that HTTP server and your AI agent.

iPhone (Camera)  <── Local Network ──>  MCP Server  <── MCP ──>  AI Agent

The AI agent never connects to the camera directly. It calls MCP tools, and the MCP server handles the rest. Four tools cover the full API:

All read-only. The agent can observe but cannot control the camera, change settings, or access recordings.

Setup Takes Two Minutes

1. Open Oriel on your camera device. In camera mode, the app shows an API token.
2. Copy the token.
3. Add the MCP server to your agent's config:

{
  "mcpServers": {
    "oriel-camera": {
      "command": "node",
      "args": ["/path/to/mcp-server/index.js"],
      "env": { "ORIEL_TOKEN": "your-token-from-step-1" }
    }
  }
}

The MCP server discovers cameras automatically using Bonjour. No IP addresses to type. No port forwarding. One token is the entire setup.

Privacy: Local Network Only

AI accessing your camera raises an obvious question: who else can see the feed?

The answer: nobody outside your network.

• The camera's HTTP server accepts connections only from the local network
• All endpoints are read-only
• Snapshots travel from phone to MCP server to agent, never leaving your WiFi
• No cloud relay, no external API calls
• You control which agents get the token. Restart the camera to revoke it.

This follows the same privacy model as Oriel's P2P streaming*. The MCP layer adds AI capabilities without adding cloud dependencies.

What AI Agents Can Do With Camera Access

Home check-ins. Ask your agent to check a specific camera while you're working. Get a description of what it sees without switching apps.

Event summaries. "What happened on the front door camera in the last hour?" The agent pulls detection events and summarizes activity patterns.

Multi-camera overview. "Are all cameras online?" The agent checks each camera's status and reports which ones are running, how long they've been active, and if any went offline.

Automated monitoring. Developers can build agents that periodically check cameras and alert on specific conditions, like a door left open after midnight.

Why MCP Instead of a Custom API

Model Context Protocol is an open standard supported by Anthropic, and adopted by a growing number of AI tools. Building on MCP means:

• Works with Claude, ChatGPT, and any other MCP client
• No custom integration code for each AI platform
• Tools are self-describing. Agents read the tool definitions and figure out how to use them.
• The open-source MCP server can be extended by anyone

A custom API would lock users into one platform. MCP keeps it open.

Open Source

The MCP server and Claude Code plugin are open source on GitHub:

OrielCamera/oriel-camera-skill

Install it in Claude Code with one command. Contributions welcome: new tools, better discovery, multi-camera workflows.

FAQ

* Oriel uses WebRTC to establish direct encrypted connections between devices. In some network environments, the connection may be relayed through an encrypted TURN server. The MCP layer operates independently on the local network and does not use WebRTC.