DEV Community

Fisher Shen (Fisher)
Fisher Shen (Fisher)

Posted on • Originally published at burn451.cloud

Give Claude access to everything you've saved: a 5-minute MCP setup for your read-later queue

#ai #claude #mcp #productivity

I save 20–30 links a week. Until recently, none of that reading was visible to the AI tools I actually think with. Claude could search the web, but it couldn't search my web — the articles I'd already decided were worth keeping.

So I built an MCP server for it. This is the setup guide: by the end, Claude (Desktop, Code, or Cursor — anything MCP-compatible) can search, quote, triage, and organize your saved articles as live context.

What you're wiring up

Burn is a read-later app with an opinion: every save gets 24 hours. Read it and it becomes a Spark; decide it's a keeper and it goes to your permanent Vault; ignore it and it burns. The point is that a reading queue should be a queue, not a landfill.

burn-mcp-server exposes that whole system to your AI agent — 26 tools across five groups:

  • Search & read — search your Vault and Sparks, pull full article content, list what's about to burn
  • Triage — let the agent read your inbox and decide what's worth keeping (with your rules)
  • Collections — create topic bundles from your saves and write synthesis overviews
  • Analysis — write structured notes back onto bookmarks
  • Auto-feed — watch an X account, RSS feed, or YouTube channel; new posts land in your queue

Step 1: Get a token

You need a free Burn account — iOS app or web.

Then: Settings → MCP Server → Generate token → Copy.

The free tier includes all 26 MCP tools (the cap is 5 saves/day; Pro at $4.99/month adds unlimited saves and the AI reading features).

Step 2: Add it to your client

Claude Desktop — edit claude_desktop_config.json (macOS: ~/Library/Application Support/Claude/):

{
  "mcpServers": {
    "burn": {
      "command": "npx",
      "args": ["burn-mcp-server"],
      "env": {
        "BURN_MCP_TOKEN": "<your-token>"
      }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Claude Code — one command:

claude mcp add burn -e BURN_MCP_TOKEN=<your-token> -- npx burn-mcp-server
Enter fullscreen mode Exit fullscreen mode

Cursor — .cursor/mcp.json in your project (or ~/.cursor/mcp.json globally):

{
  "mcpServers": {
    "burn": {
      "command": "npx",
      "args": ["burn-mcp-server"],
      "env": { "BURN_MCP_TOKEN": "<your-token>" }
    }
  }
}
Enter fullscreen mode Exit fullscreen mode

Windsurf and any other MCP-compatible client work the same way — stdio transport, one env var.

Restart the client and you should see the burn server with its tools listed. That's the whole setup.

Step 3: Ask it something you couldn't ask before

The prompts that made this click for me:

  • "Go through my Flame inbox, read each article, keep anything about AI agents and burn the rest." — triage as delegation, with the agent actually reading the content before deciding.
  • "Search my Vault for everything on context engineering and summarize the main schools of thought." — synthesis across months of your own curation.
  • "Which authors do I save most often? What does that say about my reading gaps?"
  • "I'm writing about read-later apps. Pull every relevant bookmark I have and list the claims I can support with a source." — this one is why I stopped copy-pasting URLs into chat windows.
  • "Add a watched source for Simon Willison's blog so new posts land in my Flame."

How it works under the hood

A few implementation details, since this is dev.to:

  • Auth: the server exchanges your long-lived MCP token for a short-lived Supabase session and caches it in ~/.burn/mcp-session.json — no re-login on every run. Tokens expire after 30 days.
  • Scoping: the token is scoped to your data only, enforced by Postgres Row Level Security — the server can't be talked into reading someone else's vault.
  • State machine: the triage flow (Flame → Spark → Vault, or → Ash) is enforced server-side. The agent can't "promote" an unread article straight to your permanent vault, which turns out to matter when you give an LLM write access to your knowledge base.
  • Rate limit: 30 calls/min per session.

Source is MIT-licensed: github.com/Fisher521/burn-mcp-server · npm.

Why bother?

Because "read it later" quietly became "context for later." The articles you chose to keep are the highest-signal dataset you own about what you're working on. Every other read-later tool stores that dataset; an MCP server makes it usable — your agent quotes your sources instead of hallucinating new ones.

Full docs live at burn451.cloud/mcp. If you hit a setup problem, open an issue on the repo — I read all of them.

Top comments (0)