Clipy MCP

What manovagyanik1/clipy-mcp MCP server does

The manovagyanik1/clipy-mcp MCP server exposes Clipy’s video library to MCP-compatible clients. It supports two main collections: the user’s own screen recordings and imported context videos, such as YouTube videos or local files imported through Clipy. Agents can search across both collections, locate timestamped moments, open matching points through Clipy URLs, and read the surrounding evidence.

For recordings, the server provides metadata, full timestamped transcripts, AI-generated summaries, key moments, browser diagnostics, downloadable MP4 files, and a combined markdown context bundle. Video frames can accompany relevant moments, including click markers rendered into the evidence. Browser diagnostics are privacy-redacted and identify routes, console messages, page exceptions, and failed requests without exposing headers, bodies, cookies, tokens, typed values, or raw query values.

How it works

Read operations use a Clipy recording’s public ID or its full share URL. A broad search can return hits from either the recordings library or the imported-video library, with a result type indicating which follow-up tools to use. Context documents can be read in time ranges, allowing an agent to process a section of a long video rather than loading the entire transcript at once.

The write path has two modes. The record tool launches a headless browser, captures a web application, and uploads the result to Clipy. The session tools provide a longer-lived workflow: start a session, operate the page with other browser tools, add markers or chapters, then stop and upload it. Sessions have an automatic maximum duration, and an active session can instead be discarded with abort_recording.

Markers may include page assertions or driver-reported observations. Chapters divide a recording into named sections. Transcript replacement uses a revision value from a prior read, which prevents a stale edit from overwriting a newer change.

Setup and configuration

The package is started with npx -y @clipy/mcp and uses the MCP client’s standard stdio configuration. Before launching it, authenticate with the Clipy CLI using npx @clipy/cli@latest login. The login stores credentials in the local Clipy configuration file, which the server can read without requiring the key to be copied into an MCP configuration.

For CI, containers, or a deliberately different credential, set CLIPY_API_KEY in the server process environment. The environment variable takes precedence over the local configuration file. Do not place the key in command-line arguments, where local processes may be able to inspect it.

Read tools require the recordings:read scope, which is included by default. Recording, session management, and transcript replacement additionally require the ingest scope. Headless recording also requires Playwright to be available in the server environment. Capturing the actual Mac screen is outside this MCP server’s scope and is handled by the Clipy CLI instead.

Tools and capabilities

The manovagyanik1/clipy-mcp MCP server includes tools for:

• Searching all Clipy memory or only recording titles and descriptions.
• Listing recordings and imported context documents.
• Reading recording metadata, transcripts, summaries, key moments, and compiled agent context.
• Inspecting privacy-redacted browser diagnostics.
• Waiting for transcript and summary processing to finish.
• Downloading recordings as MP4 files for local clipping or frame extraction.
• Recording headless browser sessions with viewport sweeps and login-related browser state options.
• Adding timestamped markers, named chapters, and verification evidence.
• Stopping or discarding active recording sessions.
• Reading imported videos in bounded time ranges.
• Replacing a transcript while preserving revision safety and marking the result as agent-edited.

Limitations and notes

The server does not capture the physical Mac display or a native window; its recording tools use a headless Chromium page. Write operations are unavailable to keys limited to recordings:read. Transcript and summary data may require polling before it is ready. Browser diagnostics describe page-reported evidence and intentionally omit sensitive request and input contents. Imported context videos are maintained separately from the user’s screen recordings, although the broad memory search can search both.