How to give Cursor video context: hand Cursor a screen recording over MCP (2026)

If you want to give Cursor video context today, you need a bridge: Cursor's agent reads text and images, and a screen recording is neither. Below, you set up that bridge with an MCP server in about five minutes, then take one bug from recording to fix.

All Cursor facts below come from Cursor's own pages, read on October 18, 2026.

Why Cursor cannot watch your recording

There is no place in Cursor's Agent or Chat to drop an mp4 and have the model watch it. People have asked for it: the forum request Agent and Chat should accept video as input has been open since June 2025, with users describing animations and UI flows that are hard to put into words.

What Cursor can read is text, and images. Cursor's MCP docs say an MCP server can return images, and Cursor attaches them to the chat for the model to look at. So the practical answer is to turn the recording into a transcript with timestamps and a few images from the right moments, and let the agent fetch them through MCP.

That is what Vidmatic does. Your recording is transcribed on Vidmatic's servers, every mark you draw is saved as an image, and the agent pulls both when it needs them. Nothing runs on your machine, and you do not need ffmpeg.

Give Cursor video context in three steps

1. Add Vidmatic to Cursor

Create ~/.cursor/mcp.json with this, or add the vidmatic entry to the mcpServers you already have:

{"mcpServers":{"vidmatic":{"url":"https://api.vidmatic.ai/mcp"}}}

There is no key in it. Open Cursor's MCP settings and start the sign-in for vidmatic. Your browser opens a Vidmatic page: sign in and pick what the agent may do, either Record & publish demos or Read only. Back in Cursor, the Vidmatic tools appear in the agent chat. The screen recording MCP server page has the same steps with a copy button, and the quickstart covers the rest of the setup.

Cursor's docs list a desktop sign-in callback on your own machine, and that is the one Vidmatic accepts. This is why the setup works in the Cursor desktop app and not in Cursor's web or background agents.

2. Record the bug and circle it

Install the Vidmatic Chrome extension, open the page with the bug and start recording. Talk as you would to a teammate: "The Save button stays grey after I edit the name." When you get to the problem, pick up the pen and draw on the page. Circle the button, cross out the wrong value, write the right one by hand.

Each mark is saved as an image at the moment you confirm it, together with what you were saying. That pairing is what lets the agent tell which button you meant when you said "this one".

No extension? In Vidmatic, press New Recording to record any screen or window from the browser. There is no pen in that mode, so name the element out loud and pause on it for a second. Recording your screen compares the two.

3. Ask the Cursor agent to read it

When the recording finishes, copy its link or id. In Cursor's agent chat, ask:

Read Vidmatic recording <id> with get_recording_intel. List what I asked for, look at each mark with get_annotation_image, then find the code and fix it.

The agent calls get_recording_intel and gets:

• the transcript, in short segments, each with its start and end time in milliseconds
• every mark you drew, with the moment you confirmed it and the sentence you were saying then
• a thumbnail of the recording

For each mark it calls get_annotation_image and gets the picture: the screen with your circle on it. Cursor attaches that image to the chat, so the model sees the exact button you circled. What your agent receives has the full list.

From there it is a normal Cursor session. The agent searches your code, edits the file and shows you the diff.

What it looks like on a real bug

Say a pricing page shows $240 where it should show $24, and the Upgrade button does nothing. You record 30 seconds: you say "the price is wrong" and cross out $240, then say "and this does nothing" and circle the button.

The agent gets two marks. The first comes with the sentence "the price is wrong" and an image of $240 crossed out. The second comes with "and this does nothing" and an image of the circled button. It does not have to guess which button on the page you meant. If what you said and what you drew ever disagree, trust the mark: it is tied to a position on the screen, and words like "here" are not.

Limits that are true today

• Desktop only: Cursor's web and background agents are not supported.
• Skills run through a tool call: Vidmatic also serves skills, step by step instructions such as turning a recording into one GitHub issue. In Claude Code they are slash commands. In Cursor, ask the agent to call get_skill(name="<skill>") and follow it.
• A new recording needs a moment: if the transcript is still processing, the response says how many seconds to wait, and the agent should ask again. Processing does not mean the recording is silent.
• No console or network logs: Vidmatic records what you see, say and draw. If your bug is a failed request, look at a recorder that captures browser logs; the best screen recorder for AI coding agents roundup compares the ones that do.
• Your own recordings by default: to read a recording someone shared with you, the agent needs the share token from the link.

Troubleshooting

The Vidmatic tools do not show up in Cursor

Check that the file is valid JSON and that the entry sits under mcpServers. Then finish the browser sign-in from Cursor's MCP settings.

A tool you expected is missing

The server does not push tool list changes to an open session. Reconnect vidmatic or restart Cursor.

The agent says it cannot see the recording

Make sure you signed in to the same Vidmatic account that made the recording, or include the share token if someone else made it.

Where this fits

This is the Cursor version of screen recording for AI coding agents: you show the bug once and the agent reads it as text and images. If you also use Claude Code, can Claude Code read a video covers the same gap from that side.