bookmarks-mcp

# bookmarks-mcp

**Your bookmarks, wherever they live, readable by your AI.**

Everyone saves. Browsers, read-later apps, that Pocket export from 2023 —
capture is a solved problem. Processing is not: the average bookmark is saved
once and never touched again. AI assistants are the first thing that can
actually process a pile like that — summarize, triage, turn it into a to-do
list — but they can't see it.

This is a local, read-only MCP server that fixes the "can't see it" part.

```
One command in your MCP client config. Zero accounts. Nothing is ever sent to us.
```

## Sources

| Source | How | Notes |
|---|---|---|
| Chrome, Edge, Brave, Arc, Vivaldi, Chromium | reads each profile's local `Bookmarks` file | every profile auto-detected |
| Safari (macOS) | reads `~/Library/Safari/Bookmarks.plist` | **includes your Reading List**; no extension can do this — Safari's extension API has no bookmarks surface |
| Firefox | reads `places.sqlite` per profile | needs Node ≥ 22.5 (built-in sqlite); degrades gracefully below |
| Export files | Pocket CSV, Instapaper CSV, Raindrop CSV, browser HTML, URL lists | via [parse-bookmarks](https://github.com/Fisher521/parse-bookmarks) |

Phone bookmarks: reachable once they sync to your desktop (iCloud → Safari on
Mac, Chrome sync → desktop profile). There is no way for any third-party tool
to read Safari bookmarks directly on iOS.

## Install

No account, no API key, no configuration — the server works the moment it's
installed.

Claude Code:

```bash
claude mcp add -s user bookmarks -- npx -y bookmarks-mcp
```

**Use `-s user`.** The scope defaults to `local`, which only registers the server
for the directory you ran the command in. Your bookmarks aren't tied to one
project, so register it account-wide.

Claude Desktop / Cursor / any MCP client (`mcpServers` JSON):

```json
{
  "mcpServers": {
    "bookmarks": {
      "command": "npx",
      "args": ["-y", "bookmarks-mcp"]
    }
  }
}
```

To run from source instead (for development), clone this repo next to its
dependency `parse-bookmarks` (the two directories must be siblings), run
`npm install` in `bookmarks-mcp`, and point your client at
`node /absolute/path/to/bookmarks-mcp/src/index.js`.

macOS + Safari: the app hosting the server (Claude Desktop, your terminal)
needs **Full Disk Access** (System Settings → Privacy & Security), because
`~/Library/Safari` is protected. Chromium browsers and Firefox need nothing.

## Tools

| Tool | What it does |
|---|---|
| `list_sources` | every bookmark source found on this machine |
| `list_bookmarks` | list, with folder filter + pagination |
| `search_bookmarks` | search titles / URLs / folders / tags |
| `recent_bookmarks` | "what did I save this week?" |
| `bookmark_stats` | the state of the pile: counts, age distribution, top domains |
| `parse_export_file` | parse a Pocket/Instapaper/Raindrop/HTML export from disk |
| `export_bookmarks` | everything out as Markdown, CSV, JSON, or importable browser HTML |
| `fetch_page_preview` | a few-hundred-token preview of a page (title, description, headings, excerpt) instead of the raw HTML — for when title/URL alone can't be triaged |

## Turn the pile into a to-do list

The reason this exists. Ask your assistant:

> Look at my bookmark stats, then help me triage: which of these become deep
> reads, which are tools to try, which feed a project I'm working on, and
> which should I finally let go? Give me a checklist for this week.

Or use the built-in `triage_bookmarks` prompt, which walks through exactly
that — including asking *you* to define what your saves usually become, so
the lanes are yours, not ours.

## Guarantees

- **Read-only.** This server never writes to a browser file. There is no
  delete tool. Your bookmarks cannot be harmed.
- **Zero phone-home.** No telemetry, no accounts, and not a single byte sent
  to us. The server talks to exactly two places: your local bookmark files,
  and — only when you ask it to preview a page — the bookmarked site itself,
  directly from your machine.
- **Clean output.** Tool results are data. No ads, no upsells, ever.

## Design boundaries

These are deliberate, not gaps waiting to be filled. PRs that cross them will
be declined:

- **Preview-level extraction only.** `fetch_page_preview` answers "is this
  worth my time" — it is not a reader mode and never will be a full-fidelity
  content extractor.
- **Public web pages only.** No proxies, no login walls, no special-casing
  for X/YouTube/WeChat/etc. If a page needs auth to read, this tool isn't the
  way in.
- **On-demand only.** Nothing fetches in the background, ever. Every network
  call is a direct response to a tool call you made.

## Roadmap

- Cloud sources behind user-supplied tokens (Raindrop, Readwise, …) — one
  adapter per service. PRs welcome; an adapter is a single well-defined file.
- Services that already ship their own MCP server don't need us — run both
  servers and your assistant sees everything. We only cover the places that
  have no other bridge.

## Who makes this

Built by the team behind [Burn 451](https://www.burn451.cloud), a read-later
app. This server is deliberately neutral: it reads your bookmarks for whatever
assistant and whatever workflow you choose, and it never writes to them.

## License

MIT