# Your YouTube, summarised for your AI agent.

Switch on one setting in the open-source feedql extension and the videos in your YouTube History and Watch later fill a private feed with transcripts, summaries and key moments. Your AI agent searches them through MCP or exports them to a knowledge base. Only you see them.

*By the feedql team · Output below is from our own running instance, 25 October 2026 · Published 2026-10-25 · Updated 2026-10-25*

## How do you set it up?

1. Get the extension from [github.com/feed-ql/extension](https://github.com/feed-ql/extension): download the folder and load it unpacked in Chrome (`chrome://extensions`, Developer mode, Load unpacked).
2. Click the feedql icon and sign in to your feedql account.
3. Open **Details** under YouTube to see exactly what it will do, then switch it on and allow Chrome to open `www.youtube.com`.

feedql creates a private topic called **My YouTube**. Nothing is switched on until you do it, and a later update that asks for more than you allowed pauses the site until you approve.

## What is read, and what never leaves your browser?

- **Read:** two pages, `youtube.com/feed/history` and `youtube.com/playlist?list=WL` (Watch later), fetched in your own browser with your own signed-in session.
- **Sent to feedql:** for each video its link, title, channel and which list it came from. That is all.
- **Never sent:** your cookies, your password, your subscriptions, your comments, anything from other YouTube pages.
- **Who sees it:** only you. The videos sit in your own feed and are not shared or listed anywhere; nobody else gets a copy of your list.
- **Turning it off:** stops reading at once and withdraws the access Chrome gave; what was already collected stays in your feed.

The extension is MIT-licensed and small enough to read. See its [PRIVACY.md](https://github.com/feed-ql/extension/blob/main/PRIVACY.md) and [SECURITY.md](https://github.com/feed-ql/extension/blob/main/SECURITY.md).

## What does feedql add?

- **Transcript.** For a video with captions (the author's or YouTube's automatic ones) feedql fetches the spoken words. There is no speech recognition: a video with no captions gets no transcript.
- **Summary and key points.** The summary is written from the transcript, not from the description, with up to five key moments that each link to the second they start.
- **No duplicates.** A video that feedql also found through a public source is the same item, not two.
- **Your language.** The summary is written in the language you read feedql in.

## What does the output look like?

This is a real result of the pipeline, taken from a public video our own instance collected (the format is the same for a video from your History). The summary is shown in English here; it is written in the language you read feedql in (the Ukrainian page shows the same video in Ukrainian).

| Field | Value |
| --- | --- |
| Video | OpenShell: NVIDIA's Free Agent Jail With 4 Policy Walls (Prism Labs) |
| Link | youtube.com/watch?v=JOiJnItAhNU |
| Transcript | ready, English, automatic captions |
| Summary | The video introduces NVIDIA's OpenShell, a free, open-source runtime that runs autonomous coding agents inside kernel-level sandboxes covering files, network, processes and credentials, each governed by a declarative YAML policy kept in Git. It explains the four independent security walls: Landlock-based file system confinement, allow-listed network egress, unprivileged process identities with seccomp filtering, and a credential model where the agent never holds a real secret and the runtime injects the token only when a request leaves for an approved endpoint. The video walks through the common failure modes the docs address, the one-line install and sandbox setup on Linux, macOS or WSL2 with Docker or Podman, and the project's stats including version 0.1.5, Apache 2.0 licensing, and SDKs in Python, TypeScript, Go and Rust. It closes by noting the 0.1-series caveats and how OpenShell pairs with NVIDIA's Garak vulnerability scanner. |
| 0:00 | The video opens by describing the risks of an unsandboxed coding agent on a laptop and introducing OpenShell as a fix. |
| 1:21 | It covers the project's stats, licensing, and the fact that OpenShell was built using its own agent workflows. |
| 1:41 | The four security walls are explained: file system, network, process, and credentials. |
| 3:08 | The video shows the one-line installer and sandbox setup, along with supported platforms and container requirements. |
| 5:34 | It covers production features like Kubernetes deployment, GPU support, extensibility, and observability, followed by intended users and caveats. |

## How does your AI agent use it?

Connect your agent to feedql through [MCP](https://feedql.com/llms-full.txt) (OAuth in Claude Desktop or claude.ai, or an API key). Your YouTube items are ordinary items with the source `ext:youtube`, so the same tools work.

```
search_items  query: source:ext:youtube place:watch_later
search_items  query: source:ext:youtube place:history since:7d
get_item      id: <item id from the list>
search_items  query: source:ext:youtube ~"how agents are sandboxed"
start_export  (all your feedql data, YouTube included, as a file for a knowledge base or second brain)
```

A list entry carries what the extension pushed and what feedql added. For a video it looks like this (fields trimmed, a public video):

```
{
  "title": "OpenShell: NVIDIA's Free Agent Jail With 4 Policy Walls",
  "source": "youtube", "kind": "video", "source_name": "Prism Labs",
  "url": "https://youtube.com/watch?v=JOiJnItAhNU",
  "transcript_status": "ready", "transcript_language": "en", "transcript_source": "auto",
  "key_points": [ { "t": 101, "text": "...", "url": "https://www.youtube.com/watch?v=JOiJnItAhNU&t=101s" } ]
}
```

In the sample, `t` is the second the key point starts and `text` is the point itself (left out here). For your own list the entry says `"source": "ext:youtube"` and `"place": "history"` or `"watch_later"`, and while the transcript is still being fetched `transcript_status` is `"pending"`. The meaning search (`~"..."`) finds a video by what it says, not by its title. `start_export` gives you your data as a file to load into a knowledge base.

## What are the limits?

- **Chrome only,** and the browser has to be open: the extension runs while Chrome runs.
- **Every 4 hours,** the newest 50 videos of each list. The page YouTube serves holds about the newest 90 of your History and no older ones, so a long backlog is not imported.
- **History has no watch time.** The time shown is when feedql first saw the video.
- **Transcripts take time.** feedql fetches one every 45 seconds across all videos, up to 300 a day, so a long list fills in over hours; videos with no captions stay without one.
- **Shorts and playlists** in your History are skipped.

## Frequently asked questions

**Does feedql see my YouTube password or cookies?**

No. The extension fetches the two pages in your own browser with your own session and sends feedql only the link, title, channel and list name of each video. Cookies never leave the browser.

**Who can see the videos I watched?**

Only you. They go to a private topic in your own feed; they are not shared, listed or compared with anyone else's data.

**Do I get a transcript for every video?**

Only for videos that have captions, the author's or YouTube's automatic ones. There is no speech recognition, so a video without captions has a title and a link but no summary of what is said.

**Can my AI agent read my YouTube feed?**

Yes, through feedql's MCP server or REST API with your own login or key: search source:ext:youtube, open an item for its summary and key points, or export your data.

**Is it safe to install?**

The extension is open source under the MIT license, every site starts switched off, a Details button lists the sites, pages and cookie names it uses, and an update that asks for more pauses until you approve.

## Want this for your own YouTube?

Leave your email and say you want the YouTube feed. The first seats go to a small group; the extension is open source and works with your own account.

[https://feedql.com](https://feedql.com/)

HTML version: https://feedql.com/use-cases/youtube-second-brain