Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 3 min read

Pocket died, so I built a library my AI agents can read: introducing Keepfor.me

Why I built this Pocket shut down on July 8, 2025. Omnivore is gone too. Like a lot of developers, I had years of saved links, highlights, and half-finished research scattered across tools that were either shutting dow

Why I built this

Pocket shut down on July 8, 2025. Omnivore is gone too. Like a lot of developers, I had years of saved links, highlights, and half-finished research scattered across tools that were either shutting down or locked behind a UI only I could use.

At the same time, my daily workflow changed. I spend my day with Claude and Cursor open. Those agents are good at reasoning, but they had no access to the material I had already read and saved. I would find myself re-pasting links I had captured months earlier, or watching an agent search the open web for something sitting in my own library.

So I built Keepfor.me: a personal library and AI context vault for developers and researchers. Save links, get durable copies, search everything, and let your AI agents read and save to the same library.

Capture from wherever you read

Capture was the first thing I made boring and complete. There is a Chrome and Edge extension, an iOS Shortcut, an Android share target, and the web app itself. If you are migrating, imports cover Pocket, Instapaper, Omnivore, browser bookmarks in HTML or CSV, Google Keep, Evernote, and Obsidian vaults, with dedup so running an import twice does not create a mess. Full export is available at any time. Your library is yours.

Durable snapshots, not just bookmarks

A bookmark is a pointer, and pointers rot. Every save in Keepfor.me stores a durable HTML snapshot in Cloudflare R2, so the copy survives even if the original page changes or disappears. On top of the snapshot you get reader mode parsing for a clean reading view, plus auto-tagging to keep a growing library organized without manual filing.

Hybrid search over content, notes, and tags

Search combines two layers. SQLite FTS5 handles full text search, and Vectorize adds semantic search. Both run over the article content, your notes, and your tags, so a query can find the thing you saved even when you do not remember the headline or the exact wording. In practice this means the library works the way memory works: sometimes exact, sometimes by meaning.

The MCP server is the point

The feature I built everything else around is the remote MCP server. It exposes 46 tools and works with Claude Desktop, Cursor, Windsurf, and Claude Code.

Access runs through OAuth 2.1 sign-in for remote connectors, or personal access tokens, and read and write scopes are separate. That distinction matters. You can hand an agent a read-only token and let it search your library, read highlights and notes, and answer questions grounded in sources you chose. Or you can grant write access and let it save new links into the same library while you work.

A small design detail I recommend to anyone building agent tooling: include a no-quota account tool that reports plan, trial, and quota status. When something fails, the agent can tell you whether you are out of quota instead of guessing.

Readwise launched Books with an MCP server in late October 2026, which I take as a good sign for the direction. Libraries are becoming context stores for agents, not just reading queues.

Self-hosting and stack

The core is open source under AGPL-3.0-only, at github.com/melcutz/keepfor.me. The stack is Python and FastAPI on Cloudflare Workers, with D1, R2, Vectorize, Queues, and Workers AI. Self-hosting needs a Cloudflare Workers account. There is no VPS to patch and no Docker stack to babysit. If you would rather not run anything, the hosted version is the same product with the operational side handled for you.

For context, Karakeep is the big open source incumbent in this space. Keepfor.me is a different bet: lighter, MCP and agent-first, and Cloudflare-native.

Pricing

The hosted version at app.keepfor.me is live, with signups open and a 14-day free trial, no credit card required.

Personal is $5/mo or $55/yr, with 150 AI summaries and 50 Ask questions per month. Pro is $9/mo or $90/yr, with 1,000 AI summaries and 500 Ask questions per month.

Try it and tell me what breaks

If you lost a home for your saved links when Pocket or Omnivore went away, or if you want your agents working from your own library instead of starting from zero every session, give it a try at app.keepfor.me, or self-host the open source core.

I am especially interested in feedback on the MCP tool set: which tools your agents actually use, what is missing, and how you think about read and write scopes. Comments here are welcome, and issues on GitHub are even better.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.