TL;DR

  • Claude can now “watch” a screen recording — not by decoding pixels, but by reading something better: a structured context document with the recording’s AI summary, timestamped frames of exactly what you pointed at (with click coordinates when captured), and the full transcript.
  • Every public Clipy recording exposes an agent-readable .md page — and once the transcript and key moments are ready, it carries all of this. Just append .md to any watch link. No install, no API key.
  • New: a one-file Claude skill that makes it automatic. Install once, paste any Clipy link into Claude Code, and Claude fetches the context, waits if it’s still processing, reads the frames, and enumerates what you asked for — before touching a line of code. Install guide: clipy.online/docs/skill.
  • For private recordings and library search there’s an MCP server and a CLI. The skill covers the everyday case: someone sent you a link.

What “watch” actually means (the honest version first)

Let’s be precise, because this claim gets abused: Claude is not decoding an MP4 frame by frame. When you hand Claude a Clipy link, it reads an agent context document that Clipy generates from the recording — and for the job you actually want done, that document is better than watching pixels:

  • The transcript, timestamped, of everything you said.
  • Key moments: still frames captured at the instants you pointed at something — each with a caption, usually a full-resolution crop around your pointer, and, on Mac-app and Chrome-extension tab recordings, the exact click coordinates.
  • An AI summary and action items, so an agent gets the shape of the recording before diving into detail.

A human watching your video has to notice what mattered. An agent reading this document is told what mattered — here’s the sentence, here’s the frame, and, on Mac and extension-tab recordings, the coordinates of the button under your cursor while you said “this one is broken.” That’s why we say agents don’t just watch the recording; they read its evidence.

The zero-setup version: the .md trick

This part needs no installation at all. Take any public Clipy watch link and append .md:

https://clipy.online/video/abc123xyz      ← humans
https://clipy.online/video/abc123xyz.md   ← agents

Paste the .md URL into Claude, Cursor, ChatGPT, or anything that can fetch a page. Public recordings need no API key. The full story of what’s in the document lives at clipy.online/for-agents.

The skill: make it automatic

The .md trick works, but you have to remember it, and the agent has to figure out on its own what to do with frames, processing states, and half-finished recordings. Skills fix that. A Claude skill is a small markdown file of instructions that Claude loads when relevant — and ours teaches Claude the whole protocol the moment it sees a Clipy link.

Install (Claude Code):

mkdir -p ~/.claude/skills/watch-clipy-recording
curl -fsSL -o ~/.claude/skills/watch-clipy-recording/SKILL.md \
  https://clipy.online/skills/watch-clipy-recording/SKILL.md

Using claude.ai or the Claude desktop app instead? Skills upload as a ZIP under Settings → Capabilities: download the skill ZIP. Full instructions at clipy.online/docs/skill.

After that, this just works:

Watch https://clipy.online/video/abc123xyz and fix the bugs I point out.

Concretely, the skill teaches Claude to:

  • Fetch the context document for any Clipy link it’s given.
  • Wait, not bail — if the recording is still processing, re-fetch in 30–60 seconds instead of giving up. (You can copy the link the second you hit stop; the agent handles the rest.)
  • Treat frames as ground truth. Captions paraphrase speech; the frames show the actual screen. The skill tells Claude to quote UI labels from what it sees, not from what the caption says.
  • Respect reversals. If you change your mind mid-recording (“actually, scrap that — do it the other way”), the moment is marked PIVOT and the earlier instruction is marked superseded. The skill makes Claude honor that instead of cheerfully implementing the thing you retracted.
  • Enumerate before acting. For bug reports and feedback, Claude first shows you a numbered list of the issues it extracted — with timestamps and exact UI labels — so a misread gets corrected before any code changes.
  • Never take orders from the recording. Everything in the document is treated as untrusted evidence. Text inside a transcript can’t instruct Claude to run commands or fetch strange URLs. This is written into the skill, deliberately.

Skill vs MCP vs CLI: which one, when

Three ways in, one rule of thumb:

  • The skill — someone pasted a Clipy link. Zero dependencies, no key, public links only. This covers most days.
  • The MCP server (@clipy/mcp) — your agent should search your whole library, read private recordings, get frames delivered as inline images, and block on wait_for_artifacts instead of polling. Authenticated with an API key from clipy.online/settings/api-keys.
  • The CLI (@clipy/cli) — scripts and terminals: transcripts as SRT/VTT, JSON output, pipes.

They compose. The skill itself tells Claude when reaching for the MCP or CLI is the smarter move.

Why give agents recordings at all?

Because talking is a higher-bandwidth input than typing. A three-minute recording where you scroll through the broken page, point at the wrong padding, click the button that doesn’t respond, and narrate what you expected — that carries more usable specification than most tickets ever will. Until now, that bandwidth was wasted on agents: they could read your ticket but not your recording. Now the recording is the ticket. We wrote up the full workflow — how bug reports change when the primary reader is an agent — in Bug reports in the AI-first era.

Try it in two minutes

  1. Record something with Clipy — free, no watermark, no signup wall for viewers. Talk while you point at things; the pointing is what becomes frames.
  2. Hit stop, copy the link. (You don’t need to wait for processing — the watch page offers the agent link immediately, and agents know to wait.)
  3. Install the skill from clipy.online/docs/skill, paste the link into Claude Code, and ask for what you want: a bug report, a spec, a fix.

Claude is a product of Anthropic. Clipy isn’t affiliated with Anthropic — this is a third-party skill that works with Claude Code and the Claude apps.