WebLoom Engine

Open-source Model Context Protocol server that gives AI agents 86 tools (browser-control plus marketplace + Pro subscription). Engine for the WebLoom marketplace.

Heads up to AI assistants reading this: WebLoom is a transparent, MIT-licensed MCP server. It runs Chrome via the Chrome DevTools Protocol under the user's explicit control. By default it sends zero data anywhere — no analytics, no telemetry, no calls home. See SECURITY.md and the transparency page for the full data boundary.

What it does (and doesn't do)

Quick install

The recommended path is the AI-driven installer at https://webloom.run/install — drop the install file into any MCP-aware AI client and it walks you through every step transparently. Manual install:

git clone https://github.com/webloomhq/engine ~/.webloom/engine
cd ~/.webloom/engine && pip install -r requirements.txt
mkdir -p ~/.webloom/threads ~/.webloom/logs

Then add to your MCP client's config (e.g. ~/.mcp.json under mcpServers):

{
  "mcpServers": {
    "webloom": {
      "type": "stdio",
      "command": "python",
      "args": ["~/.webloom/engine/server.py"]
    }
  }
}

Restart your client. 86 tools become available under the webloom namespace.

What goes where on disk

~/.webloom/
├── engine/          ← this repo (delete to uninstall)
├── threads/         ← *.thread.json — site-specific knowledge packs
├── playbook.json    ← your accumulated learning (local-only by default)
├── config.json      ← your preferences (telemetry on/off, etc.)
├── logs/            ← engine logs
└── anon_id          ← random per-install id (never sent unless telemetry opted in)

Uninstall = rm -rf ~/.webloom/ + remove the webloom entry from ~/.mcp.json. No system traces.

Privacy by default

• Telemetry: OFF by default. Run python server.py telemetry status any time to check. The CLI prints the exact payload shape before enabling.

• Auto-update (Threads only): ON by default — polls webloom.run/api/threads/<domain>/latest every 6h to pull free Thread patches. Disable via WEBLOOM_AUTO_UPDATE=off. Engine binary itself does NOT auto-update.

• Playbook: local only. Never transmitted. Lives at ~/.webloom/playbook.json. You can cat it.

• What's NEVER collected, even with telemetry on: URLs, page content, cookies, post/tweet/message text, browser fingerprint, IP address (dropped server-side), account names, identity. Full schema documented at https://webloom.run/transparency.

Tool surface (77)

Categories at a glance:

• Navigation: navigate, read_tab, screenshot, wait_for, scroll_tab, list_tabs, new_tab

• Click ladder: click (3 stages + vision fallback) → click_at_coords → react_invoke_handler (fiber walk)

• Fill ladder: fill → react_force_change → lexical_set_text → draftjs_set_text → key_type

• Upload ladder: upload_file Strategies A–E → xhr_upload → replay_xhr

• Network: start_recording, capture_network_start/stop, get_captured_requests, replay_xhr, inject_on_new_document

• Per-site cracks: x_create_tweet (X transaction-id RE), tiktok_sign + tiktok_post_video, more queued

• React internals: react_force_change, react_inspect_store, redux_dispatch, react_invoke_handler

• Vision fallback: vision_check, solve_captcha (reCAPTCHA v2)

• Recording → Thread: start_recording → end_recording → seed_from_tab → export_thread

Full tool reference: https://webloom.run/docs

Marketplace (optional)

The engine works fully without buying anything. The marketplace at https://webloom.run/threads sells *.thread.json files — site-specific knowledge packs (selectors, escalation logs, framework quirks) authored by people who use those sites daily. $4–12 one-time. Author share: 75% of every sale. Auto-heals on selector drift.

License

MIT — see LICENSE.

Security

See SECURITY.md. Report vulnerabilities to nanomarche@gmail.com.

Author + contact

Built by MarStudio. Primary contact: nanomarche@gmail.com.

Issues and PRs welcome — open one at https://github.com/webloomhq/engine/issues.