Video Companion

I Gave My AI Agent Access to 400+ Karakeep Bookmarks. Then This Happened

+
šŸ”„ Get the Karakeep AI Agent Skill + Companion Post Download the MCP-first Karakeep skill and view the full companion post here: https://thomasfellows.com/videos/self-hosted-karakeep-bookmark-manager-ai-agents/ GitHub for Codex, Claude Code, OpenClaw, and other agents: https://github.com/thethomasjfellows/karakeep-mcp-first-skill OpenClaw / ClawHub listing: https://clawhub.ai/thethomasjfellows/karakeep-mcp-first šŸ“Œ Turn Your Bookmarks Into an AI-Usable System In this video, I self-host Karakeep on my MacBook, import 400+ old bookmarks, connect it to OpenClaw, and test whether an AI agent can actually use my saved links inside real workflows. I cover Karakeep setup, API keys, the CLI, MCP, browser/mobile access, AI tagging, bookmark organization, highlights, and why this is much more useful than a normal bookmark manager. ā±ļø Timestamps 00:00 šŸ‘‹ Giving my AI agent access to 400+ bookmarks 01:26 šŸ› ļø Self-hosting Karakeep on a MacBook 03:31 🌐 Browser extension, API keys, and local AI features 06:58 āš™ļø API, CLI, MCP, and skills explained 11:20 šŸŒ First real test: saving and using a recipe 16:11 šŸ“„ Importing and organizing 400+ start.me bookmarks 23:11 šŸš€ A full AI workflow using Karakeep, OpenClaw, and local files 26:10 šŸ“± Using Karakeep from iPhone, Tailscale, and Telegram 28:59 🧭 Karakeep features: tags, highlights, archive, RSS, webhooks, and rules 34:04 🧠 Final thoughts: when Karakeep makes sense 🧠 Why Karakeep Is More Than a Bookmark App Karakeep is a self-hostable bookmark manager, but the real value for me is not just saving links. The value is giving my AI agent access to those links so it can search them, organize them, summarize them, tag them, highlight important parts, and use them inside actual workflows. In this walkthrough, I install Karakeep locally with Docker, connect it to OpenClaw, set up the browser extension and API key, install the CLI, add MCP support, update the Karakeep skill so it uses MCP first and CLI as a fallback, then import hundreds of old start.me bookmarks and have my agent organize them. The best part is the real workflow test: I ask my agent to research OpenClaw updates, save useful links into Karakeep, create a YouTube outline, and add a local file link back into the bookmark system. That is where Karakeep stops feeling like a bookmark app and starts feeling like part of an AI-usable knowledge system. šŸ”— Resources / Links - Updated Karakeep MCP-First AI Agent Skill on GitHub for Codex, Claude Code, OpenClaw, and other agents - https://github.com/thethomasjfellows/karakeep-mcp-first-skill - Updated Karakeep MCP-First OpenClaw Skill on ClawHub - https://clawhub.ai/thethomasjfellows/karakeep-mcp-first - Karakeep - https://karakeep.app - Karakeep Documentation - https://docs.karakeep.app - Karakeep Docker Installation Guide - https://docs.karakeep.app/Installation/docker - Karakeep API Docs - https://docs.karakeep.app/api/karakeep-api - Karakeep GitHub Repo - https://github.com/karakeep-app/karakeep - Karakeep MCP Package - https://www.npmjs.com/package/@karakeep/mcp - Karakeep MCP Container - https://github.com/orgs/karakeep-app/packages/container/package/karakeep-mcp - OpenClaw - https://openclaw.ai - Docker - https://www.docker.com - Tailscale - https://tailscale.com - start.me - https://start.me - Raindrop.io - https://raindrop.io - Ollama - https://ollama.com - Typora - https://typora.io šŸ‘¤ Who This Is For This is for people who self-host tools, use AI agents, collect a lot of links, or want their bookmarks to become part of their actual workflow instead of sitting unused in a browser folder. āœ… Key Outcomes - Self-host Karakeep locally with Docker - Connect Karakeep to AI agents through API, CLI, and MCP - Import and organize hundreds of old bookmarks - Understand when MCP is better than CLI and when CLI fallback still matters - Use Karakeep as an AI-accessible research and workflow system - Decide whether Karakeep makes sense for your own bookmark setup šŸš€ Updated Karakeep MCP-First AI Agent Skill GitHub for Codex, Claude Code, OpenClaw, and other agents: https://github.com/thethomasjfellows/karakeep-mcp-first-skill OpenClaw / ClawHub: https://clawhub.ai/thethomasjfellows/karakeep-mcp-first
Beyond The Video

Companion Updates

Key takeaways

  • Karakeep becomes much more useful when it is not just a place to store links, but a system an agent can search, organize, tag, summarize, and use inside real tasks.
  • The practical stack here is layered: the web app is for humans, the API and CLI make Karakeep scriptable, MCP gives agents a more native tool path, and the skill tells the agent which path to prefer.
  • Keeping the CLI around still matters even in an MCP-first setup because bulk inventory, fallback actions, and some feature gaps are easier to handle through terminal-accessible tooling.
  • Large bookmark imports can work, but local crawling, AI summaries, and auto-tagging can take a long time. In this case, 400+ imported bookmarks took hours to process locally.
  • The real payoff is workflow reuse: the same saved links can become research folders, video outlines, local file launchers, recipe emails, highlights, and future automation triggers.

Best suited for

This is best for people who already use AI agents and have a pile of links that never quite turn into action. It is especially useful if you are comfortable self-hosting, already use tools like OpenClaw or Codex, and want your bookmark library to become usable context instead of a drawer full of saved tabs.

If you mostly want a polished visual bookmark dashboard and do not care about automation, a hosted tool like start.me, Raindrop.io, or ordinary browser bookmarks may still be simpler.

The system Thomas built

The final setup has a few layers, and each one has a job:

  • Karakeep web app for normal browsing, saving, lists, tags, highlights, archive, RSS, webhooks, and rules.
  • Docker self-hosting so the app runs locally on the MacBook.
  • Browser extension for saving links from Chrome into the local Karakeep instance.
  • API keys for the browser extension, CLI, mobile app, and agent tooling.
  • Local AI model support for Karakeep summarization and tagging.
  • Karakeep CLI for terminal access, raw JSON, bulk work, debugging, and fallback operations.
  • Karakeep MCP server for a more AI-native interface.
  • A Karakeep agent skill that tells Codex, Claude Code, OpenClaw, and similar agents to prefer MCP first and use CLI when MCP is not enough.

That last part is the key design choice. The skill is where personal workflow rules belong. If every dessert recipe should go into a specific family recipe list, or every client login should be tagged a certain way, the skill can encode that preference.

Workflow tests from the video

The first test was intentionally small: ask the agent to find a banana bread recipe, compare results, save the best one into Karakeep, and add the ingredients to the saved item. It was not perfect, but it worked well enough to prove the point.

The bigger test was more interesting. The agent had to:

  • find the three most relevant pages about a recent OpenClaw update
  • create a Karakeep list for that research
  • save those pages into the list
  • read the pages
  • draft a YouTube outline in a local Markdown file
  • add one more Karakeep link that opens that local outline in Typora

That is where Karakeep starts to feel less like a bookmark manager and more like an AI-accessible research layer. Saved links are not just reference material anymore. They can become structured inputs inside a larger workflow.

Practical notes

For local installs, mobile access is a real decision. The iPhone app can connect through Tailscale, but if Tailscale is off, the local Karakeep instance is not reachable. For always-on mobile access, a VPS-hosted Karakeep install may be a better fit.

The import process also deserves patience. Importing 400+ start.me bookmarks worked, but the local system needed time to crawl pages, create summaries, and generate tags. After that, the agent could classify bookmarks, create lists, normalize tags, flag duplicates, and clean up the imported structure.

The MCP troubleshooting moment is worth remembering too: after installing new tools or skills, start a fresh agent session and verify the tools are actually exposed. If MCP is missing, the CLI fallback can still save the workflow while you repair the MCP config.

Resources