Skip to content
Back to search

skill-recorder

Desktop app that records your on-screen work session and uses the GitHub Copilot CLI to reconstruct it as an intent + ordered steps, then builds a reusable Skill or Automation for Microsoft Scout, Microsoft Copilot Cowork, or Copilot Studio.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today

Check it deeper on your machine

It has changed since its scan on 11 Nov 2026.

As a Claude Code plugin, install it and then ask Claude Code:

/plugin marketplace add keithmackay/mackayi/plugin install scanskillsafety@mackayiDo a deep scan of https://github.com/microsoft/skill-recorder with scanskillsafety

Or from a terminal, clone it and run it:

git clone https://github.com/keithmackay/scanskillsafety-skill ~/.claude/skills/scanskillsafetynode ~/.claude/skills/scanskillsafety/cli.cjs --deep 'https://github.com/microsoft/skill-recorder'

It checks the exact commit: every file, its packages against OSV (api.osv.dev), how it compares with this page's scan (findsafeskills.com receives the repository's address), and, if you agree, Claude reading every file with your own Claude subscription. It never runs the code.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/microsoft-skill-recorder/badge.svg)](/listings/microsoft-skill-recorder)

HTML

<a href="/listings/microsoft-skill-recorder"><img src="/api/listings/microsoft-skill-recorder/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

Install instructions from the README

As written by the repo's author. findsafeskills has not verified these steps.

Open the **[latest release](https://github.com/microsoft/skill-recorder/releases/latest)** and follow these steps:

1. On the release page, scroll to **Commit-pinned installation**. Find **Windows 11 x64 or ARM64** or **macOS or Ubuntu**, then copy the **entire command** in that section using the copy button at the top-right of the code block. 2. Open **PowerShell** from the Windows Start menu, or **Terminal** on macOS or Ubuntu. Paste the command and press **Enter**. 3. Keep that window open while installation completes. The installer downloads what it needs, builds the app on your computer, creates shortcuts, and launches Skill Recorder. The first installation can take several minutes.

> [!IMPORTANT] > **Copy the command from the release page, not a template from the technical docs.** > The release command already contains the correct version identifier; there is > nothing to replace. Do not paste commands containing > `<40-character-release-commit>` literally. > You do **not** need Git, a separate Node.js installation, or the **Manual developer > setup** instructions. The **Source code (zip)** and **Source code (tar.gz)** links > under **Assets** are not app installers.

To review the script before running it, follow the release page's **inspect-first instructions** instead. Installation remains subject to your organization's policies.

**Open it again later:** use **Skill Recorder (Source)** on your Windows desktop or Start Menu; on macOS, find it in `~/Applications`, Spotlight, or Launchpad; on Ubuntu, use the matching application entry. You do not need to reinstall each time. Advanced options, including running in the background on macOS or Ubuntu, are in [`INSTALL.md`](INSTALL.md#commit-pinned-one-line-installation).

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md · README · 362 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.

  1. # Skill Recorder
  2. **Record yourself doing a task once, then turn it into a skill your AI agent can repeat.**
  3. > [!TIP]
  4. > **Want to install Skill Recorder? [Go to the latest release](https://github.com/microsoft/skill-recorder/releases/latest).**
  5. > Scroll to **Commit-pinned installation** and copy the complete command for your
  6. > operating system. It is ready to paste: you do not need to edit it or download
  7. > the source code yourself. [Step-by-step instructions below](#install-it).
  8. Skill Recorder captures a real work session on your screen: the clicks, the app and
  9. window switches, the pages you visit, commands/output from its optional recorded terminal,
  10. and (if you want) your spoken narration. It then uses
  11. the **GitHub Copilot CLI** to reconstruct *what you actually did* as a clear **intent plus
  12. an ordered list of steps**. From there, one step turns that single run into something an
  13. agent can reuse:
  14. - a **Skill**: a `SKILL.md` procedure an agent runs on demand, or
  15. - an **Automation**: the same procedure on a schedule or trigger.
  16. Both prefer the agent's **native tools** (like the `gh` CLI or `web_fetch`) over replaying
  17. UI clicks, and generalize from your one example, so recording yourself submitting *one*
  18. form can teach the agent to submit *all* of them.
  19. <p align="center">
  20. <img src="docs/images/recorder.png" alt="Skill Recorder capture window: a record button, timer, an optional narration toggle with language and microphone settings, and readiness checks" width="420">
  21. &nbsp;&nbsp;
  22. <img src="docs/images/library.png" alt="Skill Recorder sessions view: recorded sessions on the left, the reconstructed intent and ordered steps on the right" width="520">
  23. </p>
  24. ## How it works
  25. 1. 🔴 **Record.** Hit record (or `⌘⇧R` / `Ctrl+Shift+R` from anywhere) and just do your
  26. task. Skill Recorder captures your screen and activity locally, in the background.
  27. 2. 🎛️ **Control.** While recording, a small always-on-top bar shows capture,
  28. microphone, and recorded-terminal state. Open the session-scoped terminal, mute,
  29. unmute, or switch mics, then finish or discard with confirmation.
  30. 3. 🧠 **Analyze.** Click Analyze and your selected AI provider reconstructs one overall intent and
  31. an ordered list of steps. Review and edit until it reads right.
  32. 4. ✨ **Create.** From an approved analysis, generate a reusable **Skill** and/or a
  33. scheduled **Automation**.
  34. For skills, **Review SKILL.md** optionally shows the complete, read-only file text
  35. (including frontmatter) before installation or export. Preparing a preview writes
  36. no skill files. Close and reopen it without regenerating; Add/Export from either
  37. surface uses that exact preview. Editing the plan discards the preview. You can
  38. still choose Add/Export directly without reviewing or adding a confirmation step.
  39. Review provides visibility into the generated instructions, not a security guarantee.
  40. ### Import an existing recording
  41. Instead of recording your screen, use the smaller **folder icon beside Record** to
  42. choose a local **MP4 or WebM** file, such as a downloaded Teams demonstration.
  43. Chromium checks the actual video and audio before the details screen opens.
  44. Unsupported codecs, encrypted or unsupported track layouts, corrupt media, and
  45. files outside the import limits are rejected with an explanation. No converter,
  46. standalone FFmpeg installation, or additional codec download is required.
  47. The initial limits are **512 MiB**, **15 minutes**, and **3840 x 2160 pixels**
  48. (or an equivalent pixel count). Audio must be mono or stereo at no more than
  49. 48 kHz. Import supports one video track and at most one audio track. MP4 track
  50. types recognized are AVC, HEVC, VP9, or AV1 video with AAC/Opus audio; WebM
  51. recognizes VP8, VP9, or AV1 with Opus/Vorbis. Recognized track metadata is not a
  52. codec-support guarantee: the shipped Chromium runtime must actually decode it.
  53. Import memory use is higher than the compressed file size. A Windows ARM64
  54. stress check with a 512 MiB input, 15-minute stereo audio, and the existing
  55. Whisper model resident peaked at approximately **4.6 GB (4.25 GiB)** across
  56. Electron processes. This is a measurement, not a guarantee for every file or
  57. platform, and excludes other apps and the external Copilot CLI process. The
  58. import decoder has a **2 GiB** working-set guard; recordings that exceed it are
  59. rejected even when their file size is within the limit.
  60. After validation, you can enter **Additional details**, or leave the field empty.
  61. Explain the task, supplement what is visible, or indicate where the demo happens.
  62. **Timestamps are guidance, not trimming:** other parts of the recording may still
  63. be processed. Choose the recording's spoken language when it contains audio.
  64. Analyze transcribes audio locally using the existing Whisper model (with the
  65. first-use download disclosed before proceeding); transcription failures stop the
  66. import's analysis so you can retry. Videos without audio and recordings with no
  67. detected speech are supported.
  68. The resulting activity list uses the normal review flow: edit details, reorder,
  69. add or remove activities, then create and review your skill as usual. Imported
  70. analysis uses frames, speech, and your details, not captured OS events; it cannot
  71. recover hidden clipboard contents or commands that were never visible or spoken.
  72. The original file is untouched. A local copy and derived artifacts are stored in
  73. Sessions, and deleting the session removes only that copy. Selecting a file does
  74. not send it to the cloud. Choosing Analyze sends selected frames, transcribed
  75. speech, and your details through the existing privacy-protection pipeline to
  76. your selected AI provider. Debug exports include the imported media and details, just as
  77. they include private data from live recordings. Only import recordings you are
  78. authorized to use, and keep sensitive/confidential content out of them.
  79. Live recording, its permissions and settings, and its event-first analysis are
  80. unchanged. Imports do not require microphone or screen-capture permission.
  81. For development, `npm test` includes the import contract/privacy tests.
  82. `npm run test:import:runtime` uses the reviewed Electron runtime to generate
  83. synthetic MP4/WebM files with Chromium and exercise decoding, seeking, silent and
  84. absent audio, unsupported codecs, cancellation, and source-file preservation.
  85. Run that check on each release target; it does not replace checking representative
  86. Teams files. Add `-- --large` to also import an exact 512 MiB synthetic MP4 and
  87. check rejection of a file one byte over the limit; this check needs additional
  88. memory and temporary disk space. The initial local measurements and runtime check were performed on
  89. Windows ARM64; other platforms and real meeting recordings require their own
  90. verification.
  91. ## Get started
  92. Skill Recorder is published as a **source release**: one command downloads a pinned Node.js
  93. runtime, builds the exact release commit on your machine, and adds a **Skill Recorder (Source)**
  94. app you can relaunch anytime. Nothing is installed globally. You'll need either a GitHub account with
  95. **Copilot access**, or a supported model provider's API key; the Copilot CLI ships
  96. with the app and powers both options.
  97. macOS is the primary target. Windows 11 (x64 and ARM64) is supported too (see
  98. [`WINDOWS-VALIDATION.md`](WINDOWS-VALIDATION.md)).
  99. ### Install it
  100. Open the **[latest release](https://github.com/microsoft/skill-recorder/releases/latest)** and
  101. follow these steps:
  102. 1. On the release page, scroll to **Commit-pinned installation**. Find **Windows 11
  103. x64 or ARM64** or **macOS or Ubuntu**, then copy the **entire command** in that
  104. section using the copy button at the top-right of the code block.
  105. 2. Open **PowerShell** from the Windows Start menu, or **Terminal** on macOS or
  106. Ubuntu. Paste the command and press **Enter**.
  107. 3. Keep that window open while installation completes. The installer downloads
  108. what it needs, builds the app on your computer, and creates shortcuts.
  109. **On Windows**, follow the completion notice: finish recording, choose **Quit**
  110. in every running copy, wait for them to close, then open **Skill Recorder
  111. (Source)** from the desktop or Start menu. Windows does not launch the app
  112. automatically. On macOS or Ubuntu, the installer launches Skill Recorder.
  113. The first installation can take several minutes.
  114. > [!IMPORTANT]
  115. > **Copy the command from the release page, not a template from the technical docs.**
  116. > The release command already contains the correct version identifier; there is
  117. > nothing to replace. Do not paste commands containing
  118. > `<40-character-release-commit>` literally.
  119. > You do **not** need Git, a separate Node.js installation, or the **Manual developer
  120. > setup** instructions. The **Source code (zip)** and **Source code (tar.gz)** links
  121. > under **Assets** are not app installers.
  122. To review the script before running it, follow the release page's
  123. **inspect-first instructions** instead. Installation remains subject to your
  124. organization's policies.
  125. **Open it again later:** use **Skill Recorder (Source)** on your Windows desktop or
  126. Start Menu; on macOS, find it in `~/Applications`, Spotlight, or Launchpad; on
  127. Ubuntu, use the matching application entry. You do not need to reinstall each time.
  128. Launching again reveals the existing recorder or recording controls instead of
  129. starting another copy. Older revisions without this protection must be quit
  130. manually before opening an updated revision.
  131. Advanced options, including running in the background on macOS or Ubuntu, are in
  132. [`INSTALL.md`](INSTALL.md#commit-pinned-one-line-installation).
  133. ### Then record
  134. 1. **Grant Screen Recording.** On first launch, macOS asks for Screen Recording permission;
  135. grant it and you're ready to record.
  136. 2. **Record, Analyze, Create.** Do your task, then Analyze. When no AI connection
  137. is configured, an existing GitHub Copilot login is recognized and reused
  138. automatically, including after an upgrade from 0.8.0. Otherwise, choose
  139. **Bring your own model** or **GitHub authentication** the first time AI is needed.
  140. Both options remain available to everyone, including Microsoft FTEs, through
  141. **AI connection** in Sessions. Recording needs neither.
  142. If you choose GitHub, Skill Recorder reuses a valid login or offers sign-in. Finish
  143. authorization in your browser; the app verifies sign-in and automatically retries
  144. the interrupted analysis once. There is no login terminal window to close.
  145. Canceling sign-in or leaving that analysis prevents the automatic retry.
  146. On Windows, a local Microsoft Entra tenant hint can offer a choice between
  147. **Microsoft Enterprise SSO** and a **Personal GitHub account**. This is only a
  148. routing hint, not proof of employment. Microsoft sign-in has two separate steps:
  149. finish SSO in the browser, then authorize the GitHub Copilot app. When SSO lands on
  150. `github.com/enterprises/microsoft`, close that tab, return to Skill Recorder, and
  151. choose **Authorize Copilot**. SSO alone does not authorize the app. Use the same
  152. browser profile and Microsoft-linked GitHub account for both steps.
  153. Without a hint, and on macOS/Linux,
  154. sign-in goes directly to the standard GitHub flow. No work-email prompt is required.
  155. The authorization step opens the default browser using the CLI's web OAuth flow,
  156. not a device code. Keeping the same browser profile lets GitHub reuse the session
  157. from SSO, so you can approve the GitHub Copilot app without entering a code.
  158. If GitHub asks you to sign in again, check that you are using the same browser profile.
  159. After authorization, close that tab and return to Skill Recorder; the app verifies
  160. access before continuing.
  161. Skill Recorder pins Copilot CLI 1.0.78, which supports web OAuth without a terminal,
  162. and runs `--no-auto-update login --web-flow` so sign-in does not depend on a
  163. different CLI version in the user's auto-update cache. Older bundled versions
  164. such as 1.0.71 do not support `--web-flow`.
  165. Sign-in can be canceled and times out after five minutes. If browser login fails,
  166. the app shows an error and a manual command for the bundled CLI (use PowerShell on
  167. Windows). After manual sign-in, retry the analysis yourself. Signing in from a skill
  168. or automation builder does not automatically repeat a build or installation.
  169. To inspect the script before running it, set install options, update, or uninstall, see
  170. [`INSTALL.md`](INSTALL.md).
  171. ### Bring your own model
  172. Open **AI connection** in Sessions to configure or change your connection.
  173. BYOM uses the bundled Copilot SDK/CLI without GitHub login or a Copilot
  174. subscription. Model requests go to your selected endpoint and are billed by
  175. that provider; BYOM does not mean local/offline processing.
  176. | Provider | Required fields |
  177. | --- | --- |
  178. | Microsoft Foundry / Azure OpenAI (v1) | API key, full `/openai/v1/` base URL, deployment name |
  179. | Azure OpenAI (versioned) | API key, resource host without an API path, deployment name, Azure API version |
  180. | OpenAI | API key and model ID; the standard base URL is prefilled |
  181. | Anthropic | API key and model ID; the standard base URL is prefilled |
  182. | Custom OpenAI-compatible | API key, HTTPS base URL, model ID, supported API format |
  183. Use a model supporting **image input and tool calling**. **Advanced** selects
  184. Responses versus Chat Completions for compatible providers. **Test and save**
  185. sends a small, billable synthetic-image/tool test, never a recording. Custom
  186. endpoint support depends on actual protocol and model compatibility.
  187. Keyless local services and managed-identity authentication are not exposed by
  188. this first API-key flow.
  189. Keys are protected using the operating system's key storage on Windows 11
  190. (x64/ARM64), macOS, and Ubuntu with a secure desktop keyring. If secure storage
  191. is unavailable, the key-entry screen says:
  192. > Secure space is unavailable. The Key will be saved in-memory and will be re-asked next time.
  193. There is no extra step: BYOM still works for the current app session. Non-secret
  194. settings are remembered where possible. **Forget key** removes this app's stored
  195. and in-memory credential, not the key at your provider. Finish or cancel active
  196. AI work before changing connections. Saved recordings and outputs are preserved,
  197. but live conversations are reset. Keys are not included in debug exports.
  198. BYOM connection settings and SDK conversation state are stored separately from
  199. recordings in app data. Forgetting a key does not erase conversation history.
  200. For development, `npm run test:ai` covers configuration, protected/memory-only
  201. storage, and connection lifecycle. `npm run test:ai:runtime` exercises the bundled
  202. SDK/CLI against a local synthetic provider, including tool-image transport and
  203. credential persistence checks. Neither command needs a real API key or sends
  204. requests to a paid model. Native key-storage and live-provider release checks
  205. are described in [`WINDOWS-VALIDATION.md`](WINDOWS-VALIDATION.md#ai-connection-and-credential-smoke-checks).
  206. > ⚠️ **Keep secrets out of your recordings.** Don't record, type, paste, or narrate
  207. > passwords, tokens, API keys, or other confidential info. Choosing *Analyze* sends
  208. > recording data to your selected AI provider. Skill Recorder reminds you before every recording.
  209. > Details in [What gets captured](#what-gets-captured).
  210. ---
  211. *Everything below is for people who want the details, or want to hack on the code.*
  212. ## What gets captured
  213. Recording, storage, frame extraction, and optional narration transcription all happen
  214. **on your computer**; nothing leaves while you record. Only when you choose **Analyze**
  215. does Skill Recorder send the event timeline (window/document titles, URLs, clipboard
  216. previews, and terminal commands), extracted screen images, narration text, and bounded
  217. terminal-output excerpts requested during analysis to your selected AI provider for processing.
  218. When protection is enabled, terminal commands and the complete local terminal transcript
  219. participate in the same on-device secret/PII scan before any requested excerpt is returned.
  220. The in-app "Records your screen and activity" panel spells out exactly what's collected:
  221. - **Window tracking:** active-app / window switches.
  222. - **Browser URLs:** the page you're on (macOS).
  223. - **Screen video:** recorded by Chromium; low-rate snapshots are kept only when the
  224. screen changes or a heartbeat is due.
  225. - **Clipboard:** short previews of copied text that tie steps together.
  226. - **Narration** *(optional)*: spoken commentary, transcribed **on-device** in any of
  227. Whisper's 99 supported languages (a one-time ~252 MB model download on first use).
  228. - **Recorded terminal** *(only when opened from the floating recording bar)*: command,
  229. cwd, shell, exit status, duration, and the complete terminal output. The transcript
  230. has no app-imposed size cap and counts toward the saved session's storage. It captures
  231. only the app-owned terminal for that recording—there are no global hooks, profile
  232. edits, background command monitors, or observation of existing terminals.
  233. > ⚠️ **Please don't capture secrets.** Passwords, access tokens, API keys, credentials, and
  234. > other confidential information should never be recorded, typed, pasted, shown, copied,
  235. > or narrated during a session.
  236. ## Develop from source
  237. **For developers changing the code, not for installing the app.** To use Skill
  238. Recorder, follow [Install it](#install-it) instead.
  239. Requires **Node.js 24**. After checking out a release revision:
  240. ```bash
  241. npm ci
  242. npm run compliance:licenses
  243. npm run dev
  244. ```
  245. `npm run dev` starts Vite and launches the Electron app with hot-reload; `⌘⇧R` (macOS) /
  246. `Ctrl+Shift+R` (Windows) toggles recording from anywhere. Full manual setup, the build and
  247. `dist` scripts, and the licensing boundary between local source builds and redistributable
  248. packages are in [`INSTALL.md`](INSTALL.md). Maintainers changing versions, dependencies,
  249. assets, or releases must follow [`RELEASING.md`](RELEASING.md).
  250. ## Evals
  251. The Copilot **describer** and **builders** have a fixture-based eval suite; see
  252. [`evals/README.md`](evals/README.md).
  253. ```bash
  254. npm run eval # score the describer against synthetic recordings
  255. npm run eval:terminal # score indexed terminal-output use and redaction
  256. npm run eval:builder # score the skill/automation generalization
  257. ```
  258. ## Documentation
  259. - **[INSTALL.md](INSTALL.md):** install options, inspect-first install, updating,
  260. uninstalling, and manual developer setup.
  261. - **[RELEASING.md](RELEASING.md):** maintainer release runbook.
  262. - **[evals/README.md](evals/README.md):** the describer / builder eval harness.
  263. - **[THIRD-PARTY-NOTICES.md](THIRD-PARTY-NOTICES.md):** licenses for bundled dependencies.
  264. ## Security
  265. Please don't report security vulnerabilities through public GitHub issues. See
  266. [`SECURITY.md`](SECURITY.md) for Microsoft's coordinated-disclosure process and reporting
  267. channels.
  268. ## Support
  269. File bugs and feature requests through
  270. **[GitHub Issues](https://github.com/microsoft/skill-recorder/issues)** (search existing issues
  271. first to avoid duplicates). Support is limited to the resources described in
  272. [`SUPPORT.md`](SUPPORT.md).
  273. ## License
  274. [MIT](LICENSE)
  275. ## Contributing
  276. This project welcomes contributions and suggestions. Most contributions require you to agree to a
  277. Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us
  278. the rights to use your contribution. For details, visit [Contributor License Agreements](https://cla.opensource.microsoft.com).
  279. When you submit a pull request, a CLA bot will automatically determine whether you need to provide
  280. a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions
  281. provided by the bot. You will only need to do this once across all repos using our CLA.
  282. This project has adopted the [Microsoft Open Source Code of Conduct](https://opensource.microsoft.com/codeofconduct/).
  283. For more information see the [Code of Conduct FAQ](https://opensource.microsoft.com/codeofconduct/faq/) or
  284. contact [opencode@microsoft.com](mailto:opencode@microsoft.com) with any additional questions or comments.
  285. ## Trademarks
  286. This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft
  287. trademarks or logos is subject to and must follow
  288. [Microsoft's Trademark & Brand Guidelines](https://www.microsoft.com/legal/intellectualproperty/trademarks/usage/general).
  289. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship.
  290. Any use of third-party trademarks or logos are subject to those third-party's policies.

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.