BookmarkOps

English · 繁體中文 · 简体中文

AI-assisted bookmark cleanup MCP tool for Chrome. Scan, plan, and clean up your Chrome bookmarks through any MCP-capable AI agent — Codex, Claude Code, Cursor — with destructive actions gated by your explicit dashboard approval.

Version: v1.0.0 · MCP server: @bookmarkops/mcp v0.1.3 on npm · Chrome Web Store: live.

---

Install

Recommended — Chrome Web Store: Install BookmarkOps

Reviewed and signed by Google. One click, auto-updates.

For developers — build from source: see Quick start below.

---

Core promises

• Open source under MIT — every line of code is auditable. See LICENSE.
• Runs entirely on your machine — bookmark data never leaves your browser. The only network traffic is the MCP bridge on localhost:7842, which never binds to a public interface.
• No telemetry, no analytics — BookmarkOps does not collect or transmit any usage data.
• Build it yourself — npm install && npm run build produces the exact extension you load into Chrome.

---

How it works

scan → report → plan → dryRun → preview → backup → apply → verify → restore

The workflow splits between three pieces:

1. The Chrome extension — the only thing that talks to chrome.bookmarks. Scans, applies, restores, and renders the dashboard.
2. The MCP server (@bookmarkops/mcp) — a small Node bridge launched by your AI tool over npx. Lets the AI propose cleanup plans without ever holding bookmark mutation authority.
3. You — every destructive action (apply / restore / delete backup) requires typing a confirmation phrase (APPLY / RESTORE / DELETE BACKUP) in the dashboard.

---

Quick start

1. Install the Chrome extension

For developers building from source, load as an unpacked extension:

npm install
npm run build

1. Open chrome://extensions.
2. Enable Developer mode.
3. Click Load unpacked.
4. Select the dist/ folder.

2. Wire up your AI tool

Open the BookmarkOps dashboard → AI Agent Settings → Quick Setup. Copy the JSON config and paste it into your AI tool.

Codex (~/.codex/config.toml):

[mcp_servers.bookmarkops]
command = "npx"
args = ["-y", "@bookmarkops/mcp"]

[mcp_servers.bookmarkops.env]
BOOKMARKOPS_TOKEN = "<your-token>"

Claude Code (~/.claude.json):

{
  "mcpServers": {
    "bookmarkops": {
      "command": "npx",
      "args": ["-y", "@bookmarkops/mcp"],
      "env": { "BOOKMARKOPS_TOKEN": "<your-token>" }
    }
  }
}

Cursor (.cursor/mcp.json): same shape — copy from Quick Setup.

No local install of the bridge is required — npx -y @bookmarkops/mcp downloads it on first run. The bridge requires BOOKMARKOPS_TOKEN in its environment and refuses to start without it.

Full setup walkthrough, available MCP tools, and the threat model are in src/agent/README.md.

3. Talk to your AI

"Scan my bookmarks and suggest a cleanup plan."

The agent calls scan_bookmarks and get_report, drafts a bookmark-plan.json, and queues it for your approval. Open the dashboard, review every operation, then type APPLY to execute.

---

Development

npm install
npm run dev                     # Vite dev server
npm run build                   # Production build → dist/
npm run test:task1              # Smoke test: scanner + report-generator (pure Node)
npm run test:workflow           # Smoke test: full plan workflow
npm run test:isolated-runtime   # Playwright extension runtime test (build dist/ first)
npm run validate                # Full validation suite (smoke tests + build)

For real Chrome testing, load the freshly built dist/ as an unpacked extension. The isolated runtime test uses a temporary Chrome profile and never touches your real bookmarks.

---

Permissions and safety boundary

• The manifest requests only bookmarks, storage, and alarms. No history, no external host permissions — only http://localhost:7842/* for the local MCP bridge.
• Agent Operator Mode can prepare scan / report / plan / dryRun work, but Apply, Restore, and Delete backup all require a typed confirmation phrase from you in the dashboard.
• The MCP bridge binds to 127.0.0.1:7842 only and validates a per-session token on every /enqueue call.

Full threat model: src/agent/README.md → Security model.

---

Contact

• Email: brianjhang.ai@gmail.com
• Issues & feature requests: github.com/brianjhang/bookmarkops/issues
• Personal site: brianjhang.com

I'm building BookmarkOps as a real tool I want to use. Bug reports, weird-bookmark-tree stories, and "have you considered X" emails are all welcome.

---

License

Licensed under MIT — see LICENSE. Copyright © 2026 Brian Jhang.

---

References

• MCP integration & threat model — src/agent/README.md
• Traditional Chinese quickstart — docs/zh-TW/getting-started.zh-TW.md