An MCP server built into your DVR

Your security cameras record everything. Finding the moment you care about means scrubbing through hours of CCTV footage - unless something else can do the looking for you. Agent DVR now includes a Model Context Protocol (MCP) server, the open standard AI assistants use to work with external tools, so Claude, ChatGPT, Gemini CLI or any other MCP client can browse your IP cameras, recordings and alerts the way you would, and tell you what it found.

Nothing is uploaded ahead of time and nothing is trained on your footage. When you ask a question, the assistant requests just the frames it needs from your surveillance server, looks at them, and answers.

What it looks like

Has anyone been to the front door today? Show me.
Contact sheet of four front door recordings: a courier leaving a parcel, a visitor knocking, a cat passing the parcel, the resident collecting it

Four recordings on the front door today. Looking at the frames:

  • 08:12 - a courier in a hi-vis vest left a parcel by the door and photographed it.
  • 12:40 - someone in a grey hoodie knocked, waited about 30 seconds and left. Nothing taken.
  • 15:27 - the parcel is still there; a cat walked past.
  • 17:05 - you picked up the parcel on your way in.

So one delivery, one visitor you didn't answer. Want a link to the 12:40 clip?

What birds visited the feeder this morning?
Contact sheet of four bird feeder recordings: a blue tit, a robin, two goldfinches and a wood pigeon

Between 06:30 and 09:00 the feeder camera recorded 14 visits. From the frames I can identify:

  • Blue tits - at least three separate visits, the first at 06:41.
  • A robin - 07:15 and again at 08:02, feeding from the tray rather than the tube.
  • Goldfinches - a pair at 08:30 on the nyjer feeder for nearly four minutes.
  • A wood pigeon under the feeder at 08:48 cleaning up spilled seed.

Two visits were too brief and dark to identify. The busiest period was 08:00 to 09:00.

We're heading out - arm everything with the Away profile.
Done. The server is armed with the Away profile: 8 cameras recording on detection, alerts on, the two indoor cameras switched to alerts only. I'll be here if you want to check on anything while you're out.
Which cameras are offline?
All 10 cameras are online. Garage has been reconnecting on and off since 14:20, which usually means a Wi-Fi issue rather than the camera. Everything else is streaming normally.

What the assistant actually sees

These are real replies from an Agent DVR server, exactly as an assistant receives them. Agent DVR does the work of choosing and stitching the frames, so each answer costs the assistant one small image rather than a video.

A day on one camera in a single image: one frame per recording, captioned with the time. This is what "what happened today?" fetches first.
A day on one camera in a single image: one frame per recording, captioned with the time. This is what "what happened today?" fetches first.
One recording in detail: frames sampled around the motion peak, each stamped with the time and offset, for "show me that one".
One recording in detail: frames sampled around the motion peak, each stamped with the time and offset, for "show me that one".

How it works

1

Agent DVR serves the tools

Every Agent DVR server from version 8.0.3 includes an MCP endpoint. It offers the assistant a small set of tools - list devices, take a snapshot, review a time window, read the alert log, arm and disarm - rather than raw camera access.

2

You connect an assistant

Add one address to your assistant's connector or MCP settings: your own server on your network, or ispyconnect.com/mcp to reach all your servers from anywhere. No API keys, no scripts.

3

You ask

The assistant decides which tools to use, looks at the frames it gets back, and answers in plain language - with the images, event links and times so you can check its work.

What the assistant can do

See what's happening now

Live snapshot of a garage door and driveway as returned to the assistant

Live snapshots from any camera. "Is the garage door shut?" "Who's at the gate?"

Review what happened

A whole time window as one grid of frames, one per recording, then a closer look at any event. Deliveries, visitors, animals, anything unusual.

Watch a day go by

A timeline of frames across hours or a whole day. "When did the car leave?" "What did the weather do?" "Show me the driveway over the day."

Read the alert log

Every alert with its time, camera, reason and what the AI detected. "What alerts did I have yesterday, and were any of them people?"

Summarise activity

Recordings and alerts per camera per hour, what was detected, the busiest time. "How busy was the shop today?" "When do the foxes come?"

Take control

Arm and disarm, switch profiles, enable or disable cameras, change recording modes, move PTZ cameras to presets - only when you ask, and only if you allowed control.

Two ways to connect

On your own network - free

Every Agent DVR server serves MCP on its own web address. Point an assistant running on the same computer or LAN at it and start asking. No account, no subscription.

http://localhost:8090/mcp

Claude Desktop (with our one-click extension), Claude Code, Gemini CLI, Cursor, VS Code Copilot and any other MCP client on your network.

From anywhere via iSpyConnect

Sign in once and every server on your account is reachable from claude.ai, the Claude apps, ChatGPT, Gemini CLI and anything else that lives in the cloud or on your phone. Requests travel over the same secure relay as the mobile apps: no port forwarding, and your server is never exposed to the internet.

https://www.ispyconnect.com/mcp

Included with every iSpyConnect subscription, alongside remote access and the mobile apps.

Plans from $7.95/month

Things to ask

  • Who's at the front door right now?
  • Has anyone been to the front door in the last 24 hours? Show me.
  • What birds visited the feeder this morning?
  • How busy was the driveway today, and when?
  • Show me the back garden across the day. When did it start raining?
  • What alerts did I have yesterday, and were any of them people?
  • Anything unusual overnight?
  • Which cameras are offline?
  • We're heading out - arm everything with the Away profile.
  • Send me a link to the clip of the courier.

Ready-made workflows

Assistants that support prompt templates show these as commands. Each one runs a whole multi-step check and reports back.

  • /morning_briefing What the cameras saw since last evening: activity, alerts, notable visitors, camera health.
  • /door_report Who came to a door camera, with times, what they did and a link to each clip.
  • /wildlife_report Which animals and birds a camera caught, identified as closely as the footage allows.
  • /camera_health Every server and camera checked for offline, errors and cameras that have gone quiet.
  • /day_in_review A camera across a whole day as a timeline, with a narrative of how the scene changed.

Built to be safe

  • Viewing and control are separate permissions. You grant control when you connect, and you can leave it off.
  • The assistant can never delete recordings, change camera or server settings, add or remove devices, or restart the server. Those commands are not offered at all.
  • Every request is logged on your server, so you can see exactly what the assistant asked for.
  • Nothing is stored on iSpyConnect. Frames pass through the relay to the assistant when you ask and are not kept.
  • Access from the cloud ends the moment you remove the connector from your assistant or your subscription lapses.
  • Usage is rate limited, so a misbehaving assistant cannot flood your server.

Works with

Anything that speaks MCP. These are the ones we test with:

Claude Desktop claude.ai Claude iOS & Android Claude Code Gemini CLI ChatGPT Cursor VS Code Copilot Any MCP client
Requirements: Agent DVR 8.0.3 or later. Local use is free. Cloud access needs an iSpyConnect subscription and a server connected to remote access. Assistants have their own requirements: image understanding needs a model that can see, which every current Claude, ChatGPT and Copilot model can.

Frequently asked questions

Does Agent DVR support MCP (Model Context Protocol)?

Yes. Every Agent DVR server from version 8.0.3 includes an MCP server, served at /mcp on the same address as its web interface. iSpyConnect also hosts a cloud MCP endpoint at ispyconnect.com/mcp that reaches all the servers on your account. Both expose the same tools over MCP's Streamable HTTP transport.

Can Claude see my security cameras?

Yes. Claude Desktop connects to your server with a one-click extension, Claude Code with one command, and claude.ai and the Claude phone apps through the iSpyConnect cloud endpoint. When you ask a question, Claude receives live snapshots and frames from recordings and describes what it sees.

Does it work with ChatGPT, Gemini or Cursor?

Any assistant that supports MCP works: ChatGPT through a custom connector to the cloud endpoint, Gemini CLI, Cursor, VS Code Copilot and others. The Gemini app itself does not yet let you add your own MCP servers.

Is it free?

On your own network, yes: the built-in MCP server needs no account or subscription. Reaching your cameras from the cloud or your phone needs an iSpyConnect subscription from $7.95 a month, which also includes remote access and the mobile apps.

Do I need port forwarding or to expose my NVR to the internet?

No. The local endpoint never leaves your network. The cloud endpoint uses the same secure relay as the mobile apps: your server makes an outbound connection to iSpyConnect and nothing is opened on your router.

Is my footage uploaded or used to train AI?

No. Nothing is uploaded in advance and iSpyConnect stores no video or images. Frames are fetched only when you ask and pass straight through to the assistant. What the assistant's own provider does with the images it receives is governed by that provider's policy, the same as any image you paste into a chat.

Can the assistant delete recordings or change my settings?

No. Deleting, reconfiguring, adding or removing devices and restarting are not offered to assistants at all. Viewing and control (arm, disarm, enable and disable devices, profiles, PTZ presets) are separate permissions that you grant when you connect.

Full setup instructions for every assistant are in the user guide.