Marey

What the Marey MCP server does

The Marey MCP server gives an AI agent a way to inspect changes on a screen over time. Instead of returning only a current screenshot or asking the agent to interpret a video, it captures a sequence of frames and composes them into a timestamped contact sheet. The returned image lets the agent compare states such as before an interaction, during a transition, and after a layout change.

The server is suited to visual debugging. Examples supported by the project include controls that jump during dragging, menus that open and close unexpectedly, canvases that update in the wrong order, and animations that stutter. Individual frames remain available at their original resolution when the contact sheet is not detailed enough.

How it works

An MCP client requests a capture or recording. Marey samples the selected screen area at a target frame rate, numbers and timestamps the frames, builds a contact sheet, and returns that image as MCP image content. The response also includes capture metadata and paths to the saved full-resolution frames.

Fixed-duration work uses record, while interactions whose timing is controlled by a person can use start_recording followed by stop_recording. The observe workflow records until the user finishes through a local control page. replay maintains a rolling buffer and produces clips around user markers. Only one recording can be active at a time, and an unattended open-ended recording is subject to a safety cap.

Setup and configuration

The README shows a direct installation command:

claude mcp add marey -- npx -y @anilyesilkaya/marey

Marey is distributed for execution through npx, with no runtime dependencies and no repository clone required according to the README. Capture settings can select the primary display, a virtual display, or a window. Window capture can use an exact native window ID from list_windows or a visible title substring; a relative rectangle can further limit the source area.

Recordings accept a target frame rate, an optional start delay, and presentation settings. The detail choices are overview, high, and max; callers can instead set contact-sheet columns and thumbnail width directly. Capturing a smaller window can preserve more readable detail than shrinking an entire desktop into a sheet.

Tools and capabilities

The Marey MCP server exposes ten tools:

• record captures for a specified duration and returns a timestamped contact sheet.
• start_recording and stop_recording bracket a recording controlled by the user.
• capture takes one screenshot using the same region-selection behavior.
• get_frame returns one saved frame at full resolution.
• get_frames creates a contact sheet for a time range and can apply a crop.
• list_windows reports window IDs, titles, bounds, and hidden state.
• observe records through the local control-page workflow.
• replay returns clips around markers from a rolling buffer.
• status reports the current recording state.

Raw frames are stored locally under capture directories, and recording responses identify the frame paths. A client without filesystem access can still request a selected frame through get_frame, because the frame is returned over MCP.

Limitations and notes

A contact sheet divides its available image area among its columns, so adding more thumbnail width alone does not guarantee readable text. The README recommends fewer columns, a narrower capture region, or opening an original frame when fine detail matters. The achieved frame rate and any gaps are reported rather than assumed to match the requested rate.

The server captures screen content, so it is intended for visual interaction analysis rather than DOM or application-state inspection. The README's browser demonstration requires Chrome, Chromium, or Edge and allows CHROME_BIN to specify the browser path; that requirement applies to the demonstration workflow described there.