FOR CODEX
Screen Recorder for Codex: Video Context Your Agent Can Read
Clipy is the agent-readable screen recorder for Codex. Record the task once (a bug, a flow, a walkthrough) and give Codex the share link. Through the Clipy skill convention, Codex fetches the recording's markdown document: summary, key moments with frames, click targets, cursor paths, and the timestamped transcript.
QUICK ANSWER
Clipy gives Codex video context it can read: record a bug or UI feedback once and paste the share link into the task. With the Clipy skill installed, Codex fetches the recording's .arec document with the summary, key moments with frames, click targets and the timestamped transcript. Public links need no key.
The short version
Set up Clipy for Codex in four steps
Open Codex and paste this prompt.
Setup prompt
Set up Clipy for me. It's a screen recorder you can read and record with. Read https://clipy.online/agents.md and follow its install section end to end: install @clipy/cli globally, run `clipy login`, install the Clipy skill folder for Codex with `clipy agents install codex` (it writes ~/.codex/skills/clipy/SKILL.md, or $CODEX_HOME/skills/clipy/SKILL.md when that is set), register the MCP server with `codex mcp add clipy -- npx -y @clipy/mcp@latest`, and verify with `clipy doctor --json`. Then tell me, with links, which pieces I have to install myself and what each one is for. Say plainly that your Clipy MCP tools only appear after I restart you.
It installs the CLI, writes the skill folder to ~/.codex/skills/clipy/SKILL.md, and registers the MCP server in your global Codex config.
Click Done once your agent has finished.
We cannot see your agent, so this is you telling us. It only opens the next step.
Product walkthrough. Tap for sound.
One recorder, every workflow
Show the workflow. Share the context.
See how a recording carries the request from the screen to your agent.
Explain the intent
Show the behavior and say what should change. The timestamped transcript preserves your explanation when audio is captured and processed.
Inspect the relevant moment
Key-moment frames help the agent examine the screen state. Pointer and click evidence adds detail when it is available in the capture.
Share the same evidence
A teammate watches the recording while your connected coding agent retrieves its available context. The recording's access settings still apply.
Codex runs the whole job on whatever context you gave it
Codex is at its best running a task to completion on its own, but it only ever works from the context you handed over at the start. A typed bug report is where that breaks: you write “the export button does nothing on the settings page,” Codex greps for “export,” finds four of them, and picks the wrong one. It cannot watch the screen recording you would send a teammate, because an MP4 is not a format an agent reads. So you end up writing the paragraphs you were trying to skip, and Codex still acts on your description instead of the actual screen. An agent-readable screen recorder is what closes that gap.
How it works
- 1
Record the job
Show the thing instead of writing it up. Ten seconds of pointing beats three paragraphs of describing.
- 2
Give Codex the link
The skill folder teaches Codex the convention: clipy.online/video/<id> has a fetchable .arec document. Public links need no key.
- 3
Codex ships it
Tests, fixes, docs: whatever the recording asked for, grounded in the exact frames and clicks you made.
One SKILL.md folder, then every link just works
Codex loads skills from a folder: a SKILL.md and the convention inside it. The Clipy skill is exactly that: install it once with a single command from /docs/skill, and from then on Codex knows what a clipy.online/video/<id> link is. Paste one into a task and Codex fetches the recording’s agent context on its own.
There is no per-link wiring and no key for public recordings. The link a teammate opens to watch the video is the same link Codex reads as structured text and frames. You teach the convention once; you hand over links forever.
Install once
One curl from /docs/skill drops the SKILL.md folder in. No config per project, no key per link.
Any public link resolves
clipy.online/video/<id> has a fetchable .arec document. Codex requests it the moment it sees the link.
Same link for humans
Your teammate watches the player; Codex reads the document. One URL, two readers.
What Codex reads instead of the MP4
An agent cannot watch a video. Codex reads text and images, so Clipy translates the recording into exactly those, server-side, the moment you stop. AREC carries the recording’s intent, extracted requests, suggested plan, key moments, diagnostics, capture gaps, and transcript.
Each key moment includes a timestamp, caption, extracted frame, and click coordinates when captured. So Codex does not guess which button you meant; it reads the frame and the point you clicked. The exact shape is documented at /docs/arec.
Summary, then key moments
The gist first, then the specific frames, each with a caption and, where captured, real click coordinates.
Timestamped transcript
Your narration becomes the intent: what you said, mapped to the moment you said it.
Built by us
Transcription and key-moment fusion run on Clipy’s own on-device pipeline, not a third-party API bolted on.
Is this what Codex Record & Replay does?
No, and the two fit together. Record & Replay, which OpenAI shipped for Codex in June 2026 inside the ChatGPT desktop app on macOS, watches you perform a recurring task once, such as filing an expense report, and turns that demonstration into a skill Codex can run again. It is built on Codex's computer-use capability and it is for work you will repeat.
Clipy is for the other job: something on your own product that you want changed once. A bug, a design review, a flow that should behave differently. The output is a fix or a diff, not a repeatable workflow, and what Codex needs is the evidence: the frame you pointed at, the click you made, and what you said while you did it. That is what an agent-readable recording carries, and it is the part Record & Replay was never meant to capture.
Task you will repeat
Use Record & Replay. Show Codex the workflow once and let it replay the skill next time.
Thing you want changed
Use Clipy. Record the bug or feedback and hand Codex the link with the frames and clicks inside.
Nothing to choose between
Both run in Codex. Keep the skill for the recurring task and paste a Clipy link when a fix is the goal.
Give it the link, walk away
Show the thing instead of writing it up. Ten seconds of pointing and narrating beats three paragraphs of describing, and an agent-readable screen recording lands as context Codex can act on rather than parse. That is the whole point of an autonomous run: you should be able to kick it off and leave.
Hand Codex the link and let the job finish: tests, a fix, a doc, whatever the recording asked for, grounded in the exact frames and clicks you made. Because chunks stream to the server while you record, a short clip is agent-ready within seconds of stop, under ten in the best case, so Codex starts on the real screen and not a stale description of it.
A flaky flow, fixed while you get coffee
It is late Friday and the password-reset flow is flaky. You start a recording, walk the repro once, clicking “Send reset link,” pointing at the toast that flashes green and then vanishes, saying “it says sent, but no email ever arrives,” and hit stop. The agent context is building before you have switched windows; a clip that short is ready in seconds. You start a Codex job pointed at the link and close the laptop. Codex loads the skill, fetches the context, jumps to the key moment where you clicked “Send reset link” at the captured coordinates, reads the frame with the toast, and greps for the handler. It finds the request firing before the session token is set, writes the fix and a regression test, and leaves a patch. You come back to a diff grounded in the exact frame you recorded, not the half-sentence you would have typed into the prompt.
One-time skill install
Drop the SKILL.md folder in once from /docs/skill. Every public Clipy link after that resolves to its agent context with no key and no per-link setup.
Runs the job unattended
Hand Codex the link and let it finish (tests, fixes, or docs) grounded in the frames and clicks you captured, not your description of them.
Reads the click, not a guess
Key moments carry the extracted frame plus real click coordinates on Mac-app and Chrome-extension recordings, so Codex acts on the button you actually pressed.
Unlisted by default, private in one click
Public links need no key through the skill. For private recordings or library search, point Codex at the @clipy/mcp server with an API key from clipy.online/settings/api-keys.
Handing Codex a job: description vs recording
| The old way | With an agent-readable recording | |
|---|---|---|
| What you hand over | A paragraph in the prompt and an MP4 Codex cannot open | One link that resolves to agent context Codex reads |
| Setup per task | Re-describe the repro and the UI every time | Install the skill once, then just paste the link |
| What Codex acts on | Its best guess at which element you meant | The exact frame and click coordinates you captured |
| Time to start | Codex waits on your write-up, then may pick wrong | Agent-ready in seconds, under ten in the best case for a short clip |
| Private recordings | Paste screenshots into the prompt by hand | @clipy/mcp with an API key reads your whole library |
Common questions
Does the Clipy skill work with Codex?
Yes. The skill is a plain SKILL.md folder that Codex loads like any other skill, and the convention inside it teaches Codex to resolve a clipy.online/video link to its AREC document. The underlying .arec link also works anywhere a tool can fetch a URL, so it is not tied to one agent.
How do I install the Clipy skill for Codex?
Run the one curl command from /docs/skill; it drops the SKILL.md folder in once. There is no per-project or per-link configuration after that. Every public Clipy link you paste resolves to its document automatically.
Do I need an API key for Codex to read a link?
Not for public recordings. The skill resolves those on its own with no key. Private recordings and library search need credentials: point Codex at the @clipy/mcp server, or use the CLI with a personal API key from clipy.online/settings/api-keys.
What is inside the document Codex reads?
The recording’s metadata, intent, requests, suggested plan, key moments, diagnostics when captured, capture gaps, and transcript. The exact shape is documented at /docs/arec.
Why can’t Codex just watch the recording?
Agents read text and images, not video. Clipy translates the recording into those two formats server-side the moment you stop, using its own transcription and key-moment pipeline, so Codex gets the summary, frames, and transcript instead of an MP4 it cannot open.
How fast is a recording ready for Codex?
Chunks stream to the server while you record, so the document starts building before you stop. A short clip is agent-ready within seconds of hitting stop, under ten seconds in the best case, which is what makes it the fastest agent-ready screen recorder for kicking off a job.
Do I still need Clipy if Codex has Record & Replay?
Yes, because they solve different problems. Record & Replay turns a task you repeat into a skill Codex can run again. Clipy gives Codex the evidence for a one-off change on your own product: the frame you pointed at, the click you made, and your narration, so it can open the fix. Use the skill for the recurring task and paste a Clipy link when the goal is a diff.
Does Codex get the actual click I made?
Yes. On Mac-app and Chrome-extension recordings the click coordinates are captured and paired with the frame at that instant, so Codex resolves the element you clicked instead of inferring it from a screenshot.