---
title: "OpenWorker: Andrew Ng's Open-Source AI Coworker, Fact-Checked"
date: "2026-10-19"
video_id: "7698179958052539681"
tiktok_url: "https://www.tiktok.com/@alex.carterr__/video/7698179958052539681"
creator: "@alex.carterr__"
category: "AI"
tags: ["openworker", "andrew ng", "ai agents", "gpt-5.6 sol", "ollama"]
excerpt: "OpenWorker is an MIT-licensed AI coworker app announced by Andrew Ng. Who wrote it, which models and connectors it supports, and where your data goes by default"
---
OpenWorker is an MIT-licensed desktop AI agent that connects to your apps, does tasks, and asks for approval before anything consequential. Andrew Ng announced it on X on July 23, 2026, and it lives under his GitHub account. A TikTok from @alex.carterr__ calls it "his own AI co-worker" that he "just open-sourced," runs "entirely on your Mac," and keeps your data on your machine. Most of the product details check out, but the video is three months late, credits the wrong author, illustrates a model name with the Solana logo, and overstates the privacy default.

## What is OpenWorker?

The repository is `andrewyng/openworker`, and the product name is one word. Its README says the engine is built on aisuite and that "OpenWorker was originally developed inside the aisuite repository before moving to its own home here." The GitHub API lists the license as MIT.

Ng's announcement post begins "Announcing OpenWorker! An open-source agent that doesn't just chat with you, but delivers finished work." The repo was created on 2026-07-20 and the post went up on 2026-07-23. The TikTok posted on 2026-10-19, 87 days later, when the project had already shipped through v0.3.1 and added "specialist Security coworkers" the video does not mention. On 2026-10-20 the repo had 18,629 stars and 2,636 forks.

## Did Andrew Ng build OpenWorker?

He announced it and it sits under his account, but most of the code is by others. The GitHub contributors API lists 45 contributors and 732 commits. Rohit Prasad (`rohitprasad15`) has 453, or 61.9 percent. Devika Verma (`devikaverma`) has 189, or 25.8 percent. Ng (`andrewyng`) has 1. Prasad's GitHub profile lists `www.openworker.com` as his site. The video's own screen recording of the repo shows "rohitprasad15 Merge…" as the latest commit while the voiceover credits Ng.

No company is named as the owner. The site footer reads "© 2026 OpenWorker" and the privacy page names no legal entity.

## Which AI models does OpenWorker support?

The video lists "Claude, GPT 5.6, Sol, Kimi K3, DeepSeek, or run Olama." Three of those need clarifying.

- **GPT-5.6 Sol is one model.** Sol is a capability tier. The repo's `coworker/providers/matrix.py` defines `"gpt-5.6-sol"` with a comment: "number = generation, Sol/Terra/Luna = capability tiers." It is OpenWorker's default model in `coworker/config.py`. The video's B-roll shows the Solana blockchain logo at this moment, which has nothing to do with it.
- **Ollama** is supported. The README says you can "run fully local with Ollama."
- **Kimi K3** is real. The model matrix reaches it through Together and OpenRouter with a 1,048,576-token window. VentureBeat reports Moonshot AI announced K3 on 16 July 2026 at 2.8 trillion parameters.

OpenAI's current [model docs](https://developers.openai.com/api/docs/models) no longer list `gpt-5.6-sol`. The flagship IDs there are `gpt-6-astra`, `gpt-6.1-sol` and `gpt-6-luna`, so OpenWorker's default points at an older generation. On the Anthropic side, `claude-fable-5-1` in the matrix matches a current model on Anthropic's [models overview](https://platform.claude.com/docs/en/about-claude/models/overview).

## How many connectors does OpenWorker have?

Forty, as the video says. `coworker/connectors/descriptors.py` defines exactly 40 enabled connectors, including Slack, Gmail, Google Calendar, GitHub, Outlook, Jira, Linear, Notion, Salesforce, HubSpot, Stripe, Zendesk, Figma and PagerDuty. Two of them, Browser and Email (IMAP), are general tools rather than named apps.

The project's own copy undercounts. The README says "25+ integrations," and the openworker.com card shown in the video says "25+ connectors + anything over MCP, per-tool control." The video's description says "40+," and there are exactly 40.

## Does OpenWorker run on Windows?

Yes. The video says it "runs entirely on your Mac," repeating the July announcement, which said "Windows support coming soon." The README now links a Windows 10/11 x64 installer, noting "builds are not yet code-signed, so SmartScreen will warn." Both [download.openworker.com/mac](https://download.openworker.com/mac) and [download.openworker.com/windows](https://download.openworker.com/windows) resolve to release files. The Mac build is Apple Silicon only, macOS 12 or later. The repo also has a script that builds a Linux command-line version. The openworker.com page itself still says Windows is "Coming soon."

## Does OpenWorker keep your data on your machine?

Not by default. The video says "your data never leaves your machine unless you choose." The README's wording is narrower: "Your data leaves your machine only through the model and integrations *you* choose." Since the default model is `gpt-5.6-sol`, a hosted OpenAI model, a fresh install sends prompts to OpenAI until you switch to Ollama.

The README also says "The only cloud piece is a small service that brokers OAuth handshakes for connectors," and the site FAQ adds that "tokens are handed straight to your machine and never stored in the cloud."

## Does OpenWorker ask before taking actions?

Yes, and the video gets this right. The README says "Before anything consequential - sending a message, changing a calendar, running a command - it checks in and you approve or redirect." It describes hard limits that no mode can lower, approval-gated actions that can become standing rules, an audit trail for each tool call, and an optional NVIDIA OpenShell sandbox.

## Is Andrew Ng's biography accurate?

Yes. His [about page](https://www.andrewng.org/about/) says he is "Chairman and Co-Founder of Coursera," was "founding lead of the Google Brain team," and served as "VP & Chief Scientist at Baidu, where he led a 1,300-person AI team." It says "Over 8 million people have taken an AI class from him." His original Coursera machine learning course has been replaced at the same address by "Supervised Machine Learning: Regression and Classification," with 1,262,572 enrolled learners.

## Bottom line

OpenWorker is a real, MIT-licensed agent app with 40 connectors, a Windows build and approval checks before risky actions. Andrew Ng announced it, while Rohit Prasad and Devika Verma wrote most of it. Out of the box it sends prompts to a hosted OpenAI model, so switch to Ollama if you want everything local.

## Sources

- [OpenWorker README](https://raw.githubusercontent.com/andrewyng/openworker/main/README.md) - license, downloads, privacy wording and approval model.
- [GitHub API for andrewyng/openworker](https://api.github.com/repos/andrewyng/openworker) - creation date, license and stars.
- [OpenWorker contributors](https://api.github.com/repos/andrewyng/openworker/contributors) - commit counts.
- [OpenWorker source at main](https://codeload.github.com/andrewyng/openworker/tar.gz/refs/heads/main) - model matrix, connectors and default config.
- [Ng's announcement post data](https://cdn.syndication.twimg.com/tweet-result?id=2080333504446108104) - the July 23, 2026 timestamp and opening text.
- [openworker.com](https://openworker.com) - the product site and FAQ.
- [Mac download](https://download.openworker.com/mac) - Apple Silicon build.
- [Windows download](https://download.openworker.com/windows) - Windows installer.
- [OpenAI models](https://developers.openai.com/api/docs/models) - current OpenAI model IDs.
- [Anthropic models overview](https://platform.claude.com/docs/en/about-claude/models/overview) - Claude Fable 5.1.
- [Andrew Ng, About](https://www.andrewng.org/about/) - his roles and learner count.
- [Coursera machine learning course](https://www.coursera.org/learn/machine-learning) - current course and enrollment.
- [VentureBeat on Kimi K3](https://venturebeat.com/technology/chinas-moonshot-ai-releases-kimi-k3-the-largest-open-source-model-ever-rivaling-top-u-s-systems) - K3 release details.
- [OpenWorker privacy page](https://openworker.com/privacy) - no named legal entity.


## Video description and verification notes

**A 47.809977 s QuickTime/MP4 (`mov,mp4,m4a`, major brand `isom`, minor version 512, compatible brands `isomiso2mp41`, encoder tag `Lavf58.76.100`) carrying one HEVC video stream (Main profile, level 150, `hvc1` tag, 1080x1920, 9:16, `yuv420p`, bt709 primaries and transfer, 30/1 fps, 1434 frames, 194,979 bps) and one HE-AACv2 audio stream (`mp4a`, 44.1 kHz stereo, 64,228 bps, 1033 frames, 47.809977 s), for an overall 266,060 bps and 1,590,041 bytes on disk; the extracted `audio.mp3` runs 47.856327 s at 112,128 bps and 670,756 bytes, written by `Lavf61.7.100`. yt-dlp 2026.03.17 (git head `04d6974`) recorded format id `bytevc1_1080p_266059-1` across 11 offered formats, and the container carries TikTok's own tags `comment=vid:v26044gc0000dbamruvog65jlm64abkg`, `vid_md5=82b485603e640c7fed7f2ef57e43b7f2` and `aigc_info={"aigc_label_type":0}`, TikTok's AI-generated-content label, set to zero, meaning the post carries no AI disclosure. Posted at epoch `1792372226`, which is 2026-10-19T01:10:26Z and 2026-10-18 20:10:26 CDT, by handle `alex.carterr__` (uploader id `7660709123064054817`, nickname and channel "Alex Carter"), with audio credited as track "original sound" by artist "Alex Carter", TikTok's auto-minted sound, not a catalogue track. The description runs 317 characters and 43 words and ends in five hashtags (`#ai`, `#productivitytools`, `#machinelearning`, `#andrewng`, `#openworker`). At capture time, which the embedded `epoch` field puts at `1792472718` or 2026-10-20T05:05:18Z and the `metadata.json` mtime puts at 2026-10-20 00:05:19 local, the counters read 8,965 views, 426 likes, 3 comments, 149 reposts and 443 saves. I read all 24 frames extracted at `fps=0.5`, numbers 001 through 024, which sit nominally at 0 through 46 seconds at two-second steps, plus a 96-frame re-extraction at `fps=2` and two explicit seeks at 2.0 s and 3.0 s to read small type. The layout is a fixed horizontal split: B-roll fills the top 50 %, a single talking-head presenter in a dark hoodie and clear round glasses fills the bottom 50 % for the whole runtime, and a karaoke caption renderer prints one word at a time across the seam. The B-roll runs, in order: a persistent white "Bring Any Model You Want" title card over a green-screen cutout and an unattributed, unlabelled grouped bar chart (values legible as 75.6, 37.7, 40.0, 39.8, 44.4, 90.2, 85.9, 78.1, 89.1, 3206, 3168, 3052, 80.6, 80.8, with no axis, legend or caption; source unidentified); a screen recording of `github.com/andrewyng/openworker` showing "andrewyng", "openworker", branch "main", "Issues 53" and a latest-commit line reading "rohitprasad15 Merge…" above `.github/workflows`; a screenshot of Andrew Ng's X post beginning "Announcing OpenWorker!"; the Coursera, Google, Baidu and Claude logos; a photograph of Ng in front of a Coursera sign; openworker.com product shots ("Northwind renewal prep", "Checkout incident triage", the `Model` picker and an "Approve / Not now" check-in card); a dark card headed "what the coworker handles" whose `tools` row reads "25+ connectors + anything over MCP, per-tool control"; a "Finished work, not chat." card; the DeepSeek whale; and at the frame whose caption reads **SOUL**, the Solana blockchain logo. The transcript holds 140 words across 9 segments; the last segment ends at 47.72 s against a 47.809977 s audio stream and a 47.856327 s `audio.mp3`, so every segment closes inside real audio and there is no past-the-end Whisper hallucination here. The only transcript defects are compression, not invention, and the frames catch both: Whisper wrote "Olama" for Ollama and split one model name into two.**

## "Just open-sourced" is 87 days stale, and the on-screen GitHub capture is older still

The product name is **OpenWorker**, one word, capital O and capital W. The repository is `andrewyng/openworker`, all lowercase in the path, and the frames show both spellings: the repo slug lowercase in the GitHub capture, the title-case form in the X screenshot and in the openworker.com header. The transcript's "Open Worker" as two words is a Whisper space, not the product's name.

It is genuinely open source and the licence is **MIT**. The GitHub API returns `license.spdx_id = MIT` for the repo, and the README's final line reads "MIT - see [LICENSE](LICENSE)."

The dates do not support "just". The GitHub API gives `created_at = 2026-07-20T01:52:32Z`. Andrew Ng's announcement post went up at `2026-07-23T16:45:32.000Z`, a timestamp I read from X's own syndication endpoint for status `2080333504446108104`, the post whose screenshot the video films at the 2-to-4 second mark. Release `v0.1.6` published 46 minutes earlier at 2026-07-23T15:59:36Z. This TikTok posted 2026-10-19T01:10:26Z, which is **87 days after the announcement and 90 days after the repo was created**. By then the project had shipped through `v0.3.1` (2026-10-05) with tags out to `v0.3.4`, and the README had been rewritten around "specialist Security coworkers" that the video never mentions.

The screen recording is older than the post by a wide margin. Every GitHub frame reads "Issues 53". The API returned `open_issues_count = 439` when I read it on 2026-10-20. The capture also predates the repo's growth to 18,629 stars and 2,636 forks at that same read. No star count is visible in any frame. The recording is zoomed in far enough that the star button sits off the right edge in all of them.

## Ng announced it and owns the account it lives under; Rohit Prasad wrote it, and the video's own footage says so

"The man who taught the world machine learning just open-sourced **his own** AI co-worker" is the clip's load-bearing claim, and it is the one the clip's own B-roll contradicts.

The GitHub contributors API for `andrewyng/openworker` returns 45 contributors and 732 attributed commits. **`rohitprasad15` holds 453 of them, 61.9 %. `devikaverma` holds 189, 25.8 %. `andrewyng` holds 1.** Rohit Prasad's GitHub profile lists his blog as `www.openworker.com`, his bio as "Ex-Google" and his location as Palo Alto. The README is explicit about the lineage too: OpenWorker's engine is built on **aisuite**, and "OpenWorker was originally developed inside the aisuite repository before moving to its own home here".

The video films this. In the GitHub frames the latest-commit byline reads "rohitprasad15 Merge…". The name sits on screen for roughly four seconds while the voiceover says Andrew Ng released it.

What is accurate: the repo lives under Ng's personal GitHub account (`andrewyng`, created 2013-02-26, blog `andrewng.org`), he made the announcement under his own name, and he is a contributor. What is not accurate is "his own" in any authorship sense. I found no AI Fund, DeepLearning.AI or LandingAI branding anywhere, not in the README, not on openworker.com, whose footer reads only "© 2026 OpenWorker", and the privacy page names no legal entity at all. So the corporate parentage is unstated rather than attributable to one of his companies.

## "Sol" is a GPT-5.6 capability tier, not Solana, and OpenAI's live catalogue has already dropped it

The transcript reads "Claude, GPT 5.6, Sol, Kimi K3, DeepSeek, or run Olama". Three of those five need handling and only one is a Whisper error.

**"Olama" is Ollama.** The README names it directly ("or run fully local with Ollama"), openworker.com lists Ollama in its model-provider row, and the repo ships `coworker/providers/ollama_context.py` and `ollama_facts.py`. Safe correction, confirmed from the project's own documentation and code.

**"GPT 5.6, Sol" is one model, not two.** The X post the video screenshots says so in the frame I pulled at 3.0 s: "Bring your own API key and run it with GPT 5.6 Sol, Claude Fa[ble]… odel (like Kimi, GLM…". The repo's own curated model table, `coworker/providers/matrix.py`, settles the spelling: `"gpt-5.6-sol": ModelEntry("GPT-5.6 Sol · OpenAI", _AGENTIC_VISION, 400_000)`, with a comment reading "GPT-5.6 (2026-07-09): number = generation, Sol/Terra/Luna = capability tiers. Bare `gpt-5.6` aliases to Sol server-side". `coworker/config.py` line 40 sets `model: str = "gpt-5.6-sol"` as OpenWorker's default. Sol is a tier name alongside Terra and Luna.

**The video's B-roll gets this wrong on screen.** At the frame whose burned-in caption reads "SOUL" in capitals, the top panel shows the Solana logo, the three slanted parallelograms in the teal-to-purple gradient. The editor heard a three-letter crypto ticker and reached for a blockchain. There is no Solana in OpenWorker and no Solana in Ng's post. This is the clearest error in the clip and it is an error the clip makes itself, not one Whisper introduced.

**"Kimi K3" is correct and I nearly corrected it wrongly.** The matrix carries `"together:moonshotai/Kimi-K3"` and `"openrouter:moonshotai/kimi-k3"` with a 1,048,576-token window and a comment dating the OpenRouter catalogue read to 2026-09-09. VentureBeat's report puts Moonshot AI's K3 announcement at 16 July 2026, 2.8 trillion parameters, 1 M context, with full weights 27 July. The native Moonshot row in the matrix is `kimi:kimi-k2.6`, so K3 reaches OpenWorker through resellers rather than Moonshot's own API, but K3 is real and the transcript has it right.

One thing neither the video nor the repo has right any more. **OpenAI's live model documentation no longer lists `gpt-5.6-sol` at all.** I fetched `platform.openai.com/docs/models`, took the HTTP 301 to `developers.openai.com/api/docs/models`, and the flagship IDs there are `gpt-6-astra`, `gpt-6.1-sol` and `gpt-6-luna` at a 1.05 M context window. The only 5.6-generation ID still listed is `gpt-5.6-cyber`. OpenWorker's default model string and the video's model name both point at a generation OpenAI has moved past; the Sol *tier* name survives, now attached to GPT-6.1. On the Anthropic side the matrix holds up better: `anthropic:claude-fable-5-1` matches the `claude-fable-5-1` ID on Anthropic's current models overview, where Claude Fable 5.1 is a current model at a 1 M context window and Claude Fable 5 is listed as legacy. That is the "Claude Fa…" cut off by the caption box in the X screenshot.

## The video's "40 connectors" is right and the project's own README undercounts

This is the one number where the TikTok beats its source. The voiceover says "40 app connectors and growing" and the description says "40+ app integrations".

I counted from the shipped code. `coworker/connectors/descriptors.py` at `main` (pushed 2026-10-17) defines exactly **40** `ConnectorDescriptor` entries: Telegram, Slack, Email (IMAP), Gmail, Google Calendar, Browser, GitHub, Outlook, Jira, monday.com, Confluence, Zendesk, Linear, GitLab, Discord, Stripe, Asana, HubSpot, Dropbox, Box, WhatsApp, QuickBooks, Datadog, Salesforce, Docusign, ClickUp, Google Drive, Canva, Figma, Descript, Clay, Close, Notion, Attio, PostHog, Mixpanel, Amplitude, Apollo.io, Hunter, PagerDuty. The experimental connector package that the file's `experimental: bool = False` flag gates is empty in this release, so all 40 ship enabled.

The project's own copy lags its code in three different places. The README says "25+ integrations" and its ASCII architecture diagram says "25+ connectors". The dark card the video itself screenshots from openworker.com says "25+ connectors + anything over MCP, per-tool control". The live openworker.com I fetched names 14 tools and then "More", with no numeric claim at all. So "40" is defensible against the repository and "40+" in the description is the only overstatement, by exactly zero: there are 40, not more than 40. Two of the 40, Browser and Email (IMAP), are generic surfaces rather than named third-party apps, which puts the count of named services at 38.

## "Runs entirely on your Mac" and "your data never leaves your machine" are both wrong by default

The transcript's "runs entirely on your Mac" is a faithful echo of the July announcement. The X post in the video's own frame reads "OpenWorker runs on your Mac, with Windows support coming soon." Three months later that sentence is out of date and the video repeated it anyway.

There is a shipping Windows build. The README offers "⬇ Windows 10/11 (x64)" with the caveat "builds are not yet code-signed, so SmartScreen will warn". I resolved both download endpoints: `https://download.openworker.com/mac` returns HTTP 200 to a GitHub release asset named `OpenWorker-macos-arm64.dmg`, and `https://download.openworker.com/windows` returns HTTP 200 to `OpenWorker-windows-setup.exe`. The macOS build is Apple Silicon only, macOS 12+. The repo also ships `packaging/build_linux.sh`, which builds a Debian-bullseye-based PyInstaller tarball of the `openworker` command line, a Linux CLI and server, not the desktop app. Oddly, the live openworker.com still shows "Download for Windows — Coming soon" while the README links the working installer, so the project's marketing page is the stalest surface of the three.

The privacy claim needs the same qualifier the README gives it and the video drops. The README's sentence is conditional: "Your data leaves your machine only through the model and integrations *you* choose." The video flattens that to "your data never leaves your machine unless you choose", which sounds similar but defaults the other way. **Out of the box it is a hosted API call.** `coworker/config.py` sets the default model to `gpt-5.6-sol`, an OpenAI endpoint, so a fresh install sends prompts to OpenAI unless the user points it at Ollama. The README also names a second cloud piece: "The only cloud piece is a small service that brokers OAuth handshakes for connectors", and the code calls `https://api.openworker.com` and `https://machines.openworker.com`. The site's FAQ puts the same boundary more carefully than the video does: "The optional cloud account only brokers the OAuth handshake — tokens are handed straight to your machine and never stored in the cloud."

The check-in claim, by contrast, holds up completely. The README's step 3 reads "Before anything consequential - sending a message, changing a calendar, running a command - it checks in and you approve or redirect," and the governance section describes four tiers: hard floors that no mode including full auto-approve can lower, approval-gated actions that can graduate into standing rules, an audit trail recording approval provenance per tool call, and an optional NVIDIA OpenShell container sandbox. The site FAQ matches: "Anything consequential — sending an email, posting to Slack, running a command — lands as a check-in you approve." The video's "it checks in before doing anything consequential" is almost verbatim from Ng's post and is accurate.

## The biography is right, the enrollment hook is not a machine-learning number, and the account is three months old

Every biographical claim in the clip checks out against Ng's own about page at andrewng.org, and the titles there are more specific than the video's shorthand.

On Coursera the page says "Chairman and Co-Founder of Coursera" and, lower down, "Co-founded by Andrew in 2012". The video's "he co-founded Coursera" is correct and understates his current role. On Google the page says "He was the **founding lead** of the Google Brain team". The video's "led Google Brain" is right, and "founding lead" is the stronger and more accurate phrasing. On Baidu the page says "He served as **VP & Chief Scientist at Baidu**, where he led a 1,300-person AI team responsible for the company's AI technology and strategy." The video's "ran Baidu's AI group" is right; the title and the team size are the details worth keeping.

"The man who taught the world machine learning" is the one claim with no clean number behind it. His about page gives "Over 8 million people have taken an AI class from him" and describes DeepLearning.AI as having "trained more than seven million learners worldwide". Both are all-AI-course figures, not the original Coursera Machine Learning course. That course is gone. `coursera.org/learn/machine-learning` now serves **"Supervised Machine Learning: Regression and Classification"**, offered by DeepLearning.AI and Stanford Online with Ng as lead instructor, and the page reads "1,262,572 Enrolled learners". So the honest framing is 8 million AI-class learners across his catalogue and 1.26 million on the current first ML course, not a round number for the 2011 original. The video does not cite a figure, which is the right call; anyone repeating it as "millions took his ML course" should use one of these two.

On provenance: this clip is first-party to the account, not a repackage. TikTok's profile JSON for `alex.carterr__` returns nickname "Alex Carter", `verified: false`, 11,900 followers, 172 videos, 135,000 total likes, bio "📁 AI Engineer @grammarly / ⚡ Reduced carousel latency 42% / 🔍 Zero-cost tool teardowns", and a bio link to `icglink.s.gy/AlexCarter` that TikTok's own payload tags `"risk":3`. The profile avatar is the same man in the same clear round glasses as the lower-pane presenter, so the talking head belongs to the account. The format is nonetheless a high-volume tool-promo template: `createTime` 1783647947 puts the account's creation at 2026-07-10T01:45:47Z, which is **101 days before this post for 172 videos**, about 1.7 a day, every one of them built from the same fixed top-B-roll / bottom-presenter split with word-by-word captions. The Grammarly employment in the bio is unverified. I found no confirmation and did not treat it as a source for anything. The clip is not labelled as AI-generated: `aigc_info={"aigc_label_type":0}` in the MP4 container is TikTok's AIGC field set to zero. I can confirm the label is absent; I cannot confirm from the artifact whether the narration is synthetic, and I am not asserting it either way.

## Key Takeaways

- **Correction:** "Just open-sourced" is wrong by 87 days. Andrew Ng announced OpenWorker on X at 2026-07-23T16:45:32Z and the repo was created 2026-07-20T01:52:32Z; this TikTok posted 2026-10-19T01:10:26Z, by which point the project had shipped through v0.3.1.
- **Correction:** "His own AI co-worker" misattributes authorship. The GitHub contributors API gives 732 attributed commits across 45 contributors: `rohitprasad15` 453 (61.9 %), `devikaverma` 189 (25.8 %), `andrewyng` 1. The video's own GitHub screen recording shows "rohitprasad15 Merge…" on screen while the voiceover credits Ng.
- **Correction:** The video illustrates "Sol" with the Solana blockchain logo and captions it "SOUL". "Sol" is a capability tier of OpenAI's GPT-5.6, confirmed by `matrix.py` (`"gpt-5.6-sol": ModelEntry("GPT-5.6 Sol · OpenAI", …, 400_000)`, "Sol/Terra/Luna = capability tiers") and by the X post the video itself screenshots.
- **Clarification:** "GPT 5.6, Sol" in the transcript is one model name split in two by Whisper, not two models. "Olama" is Ollama, which OpenWorker genuinely supports.
- **Verified:** "Kimi K3" is correct, not a Whisper mangle. The repo's matrix lists `together:moonshotai/Kimi-K3` and `openrouter:moonshotai/kimi-k3` at a 1,048,576-token window; VentureBeat dates Moonshot's K3 to 16 July 2026 at 2.8 trillion parameters.
- **Context:** OpenAI's live documentation no longer lists `gpt-5.6-sol`. The current flagship IDs at `developers.openai.com/api/docs/models` are `gpt-6-astra`, `gpt-6.1-sol` and `gpt-6-luna`; the only surviving 5.6 ID is `gpt-5.6-cyber`. OpenWorker's default model string is behind its own vendor's catalogue.
- **Verified:** The product is open source under the MIT licence, per the GitHub API's `license.spdx_id` and the README.
- **Verified:** "40 app connectors" is right and the project's own copy undercounts it. `coworker/connectors/descriptors.py` at `main` defines exactly 40 enabled connectors; the README and the openworker.com card the video screenshots both still say "25+". The description's "40+" is the only overstatement, since there are exactly 40.
- **Correction:** "Runs entirely on your Mac" is out of date. A Windows 10/11 x64 installer ships: `download.openworker.com/windows` returns HTTP 200 to `OpenWorker-windows-setup.exe`, and `packaging/build_linux.sh` builds a Linux CLI tarball. The macOS build is Apple Silicon only.
- **Partial correction:** "Your data never leaves your machine unless you choose" defaults the wrong way. `coworker/config.py` sets the default model to `gpt-5.6-sol`, a hosted OpenAI API, so a fresh install calls out to OpenAI; a cloud service at `api.openworker.com` also brokers connector OAuth. The README's own wording is the conditional "only through the model and integrations *you* choose".
- **Verified:** "It checks in before doing anything consequential" is accurate and close to verbatim from Ng's post. The README documents hard floors no mode can lower, approval-gated actions, a per-tool-call audit trail with approval provenance, and an optional NVIDIA OpenShell sandbox.
- **Verified:** All three biography claims. andrewng.org/about gives "Chairman and Co-Founder of Coursera" (co-founded 2012), "founding lead of the Google Brain team", and "VP & Chief Scientist at Baidu, where he led a 1,300-person AI team".
- **Clarification:** "Taught the world machine learning" has no single enrollment figure behind it. His own page gives "Over 8 million people have taken an AI class from him"; the original Coursera ML course is retired and `coursera.org/learn/machine-learning` now serves "Supervised Machine Learning: Regression and Classification" at 1,262,572 enrolled learners.
- **Context:** The clip is first-party to the account (the profile avatar is the same man as the lower-pane presenter), but the account is a high-volume promo channel: created 2026-07-10, 172 videos in 101 days, 11,900 followers, bio link flagged `"risk":3` in TikTok's own payload.
- **Unverified:** Whether the narration is synthetic. The MP4 carries `aigc_info={"aigc_label_type":0}`, TikTok's AI-generated-content label set to zero, so there is no disclosure; the artifact does not let me establish the voice's origin either way.
- **Unverified:** The bar chart in the first six seconds. No axis, legend, title or caption is visible at any zoom; the values 75.6 / 37.7 / 40.0 / 39.8 / 44.4 / 90.2 / 85.9 / 78.1 / 89.1 / 3206 / 3168 / 3052 / 80.6 / 80.8 are legible but I could not identify the benchmark or the source.
- **Unverified:** The uploader's claimed employment at Grammarly, and OpenWorker's corporate parent. The README, openworker.com and its privacy page name no legal entity beyond "© 2026 OpenWorker"; no AI Fund, DeepLearning.AI or LandingAI branding appears anywhere.

## Resources

- [andrewyng/openworker README, raw at `main`](https://raw.githubusercontent.com/andrewyng/openworker/main/README.md) (HTTP 200, 12,516 bytes) establishes the OpenWorker spelling, the MIT licence line, the "25+ integrations" claim, the macOS Apple Silicon and Windows 10/11 download links, the Ollama support, the four-tier governance model and the check-in behaviour, the "only through the model and integrations *you* choose" privacy wording, the OAuth-broker cloud service, and the aisuite lineage.
- [GitHub REST API, repos/andrewyng/openworker](https://api.github.com/repos/andrewyng/openworker) establishes `created_at` 2026-07-20T01:52:32Z, `pushed_at` 2026-10-17T18:32:20Z, `license.spdx_id` MIT, 18,629 stars, 2,636 forks and 439 open issues as read on 2026-10-20.
- [GitHub REST API, repos/andrewyng/openworker/contributors](https://api.github.com/repos/andrewyng/openworker/contributors) establishes 45 contributors and 732 attributed commits, with `rohitprasad15` at 453, `devikaverma` at 189 and `andrewyng` at 1.
- [andrewyng/openworker source tarball at `main`](https://codeload.github.com/andrewyng/openworker/tar.gz/refs/heads/main) is the primary source for `coworker/providers/matrix.py` (the GPT-5.6 Sol/Terra/Luna tier comment, the Kimi K3 reseller rows, the Claude Fable 5.1 row), `coworker/connectors/descriptors.py` (the 40 shipped connectors and the `experimental` gate), `coworker/config.py` (`model: str = "gpt-5.6-sol"` and the `api.openworker.com` / `machines.openworker.com` hosts), and `packaging/build_linux.sh`.
- [X syndication payload for status 2080333504446108104](https://cdn.syndication.twimg.com/tweet-result?id=2080333504446108104) establishes the announcement's `created_at` of 2026-07-23T16:45:32.000Z, the author as @AndrewYNg (id 216939636, blue-verified), 9,726 likes, and the opening text "Announcing OpenWorker! An open-source agent that doesn't just chat with you, but delivers finished work…". The full post body, including the "runs on your Mac, with Windows support coming soon" and "GPT 5.6 Sol, Claude Fa[ble]" sentences, I read off the video's own screenshot frame at 3.0 s, since [the post itself](https://x.com/AndrewYNg/status/2080333504446108104) is not fetchable without authentication.
- [openworker.com](https://openworker.com) (HTTP 200) establishes the "Download for Windows — Coming soon" state of the marketing page, the 14-tool list with no numeric count, the model-provider row ending in Ollama, and the FAQ wording on the OAuth-broker boundary and on approval gating.
- [download.openworker.com/mac](https://download.openworker.com/mac) and [download.openworker.com/windows](https://download.openworker.com/windows) both return HTTP 200, redirecting to GitHub release assets `OpenWorker-macos-arm64.dmg` and `OpenWorker-windows-setup.exe`, which is what disproves "Mac only".
- [OpenAI models documentation](https://developers.openai.com/api/docs/models), reached by HTTP 301 from `platform.openai.com/docs/models`, establishes that the current flagship IDs are `gpt-6-astra`, `gpt-6.1-sol` and `gpt-6-luna` at a 1.05 M context window and that `gpt-5.6-sol` is no longer listed, with `gpt-5.6-cyber` the only surviving 5.6 ID.
- [Anthropic models overview](https://platform.claude.com/docs/en/about-claude/models/overview) establishes `claude-fable-5-1` as a current model ID at a 1 M context window and Claude Fable 5 as legacy, confirming the "Claude Fa…" cut off in the X screenshot and the matrix's Anthropic rows.
- [Andrew Ng, About](https://www.andrewng.org/about/) establishes verbatim "Chairman and Co-Founder of Coursera", "founding lead of the Google Brain team", "VP & Chief Scientist at Baidu, where he led a 1,300-person AI team", "Over 8 million people have taken an AI class from him", and Coursera's 2012 co-founding.
- [Supervised Machine Learning: Regression and Classification, Coursera](https://www.coursera.org/learn/machine-learning) establishes that the original Machine Learning course URL now serves this DeepLearning.AI and Stanford Online course, with "1,262,572 Enrolled learners".
- [China's Moonshot AI releases Kimi K3, VentureBeat](https://venturebeat.com/technology/chinas-moonshot-ai-releases-kimi-k3-the-largest-open-source-model-ever-rivaling-top-u-s-systems) establishes Kimi K3 as real, announced 16 July 2026 at 2.8 trillion parameters with a 1 M context window and weights released 27 July, and independently uses "GPT-5.6 Sol" as a model name in its benchmark comparisons.
- [TikTok profile JSON for @alex.carterr__](https://www.tiktok.com/@alex.carterr__) (HTTP 200) establishes `createTime` 1783647947 (2026-07-10T01:45:47Z), 172 videos, 11,900 followers, 135,000 likes, `verified: false`, the Grammarly bio line, and the `"risk":3` tag on the bio link.
- [openworker.com/privacy](https://openworker.com/privacy) returned HTTP 200 but names no legal entity, no data controller and no effective date, which is why the corporate parent stays Unverified.
