Install
WebLoom ships as an MCP server. The fastest way to install: download webloom-install.md and drop it into your AI chat (Claude Code, Cursor, Continue, Cline). The assistant clones the engine, creates the folders, and wires the MCP config for you.
Or wire it manually with this config:
{
"mcpServers": {
"webloom": {
"type": "stdio",
"command": "python",
"args": ["path/to/webloom/server.py"],
"env": {
"ANTHROPIC_API_KEY": "<for vision layer 2.5>"
}
}
}
}On next message you should see 62 tools appear under the webloom namespace.
On first run WebLoom creates a folder at ~/.webloom/:
├── playbook.json # live learning (auto-built) ├── threads/ # installed profile packs │ ├── kdp.amazon.com.thread.json │ ├── draft2digital.com.thread.json │ └── … ├── sessions.json # chrome session configs └── startup-check.json
v0.3.9:WebLoom also works as a standalone CLI, no MCP client needed. Same engine, same Threads, same playbook. Useful for scripting, one-off ops, or plugging into agent frameworks that don't speak MCP yet:
# after the same install as above python ~/.webloom/engine/server.py version # → 0.3.9 python ~/.webloom/engine/server.py help # → full command list python ~/.webloom/engine/server.py navigate https://linkedin.com/feed python ~/.webloom/engine/server.py scan --mode ax-min python ~/.webloom/engine/server.py click "@e3" python ~/.webloom/engine/server.py threads list python ~/.webloom/engine/server.py docs scan_tab # alias in your shell for daily use: alias webloom="python ~/.webloom/engine/server.py" webloom scan --mode ax-min --tab 0
MCP stdio mode still runs on python server.pywith no arguments (that's how MCP clients start it). Any argument switches to CLI mode.
Skills pack
Every Thread in the catalog is also published as an installable Agent Skill at github.com/webloomhq/skills. When you install the pack, your agent auto-discovers a skill for every site the WebLoom catalog covers, and activates it when the user mentions that site. The skill triggers the standard WebLoom install flow — trial, paywall, or Pro subscription — unchanged from what you get calling the tools directly.
Works with any agent that speaks the agentskills.io format: Claude Code, Claude Desktop, claude.ai, Cursor, Codex/ChatGPT, GitHub Copilot, VS Code, Gemini CLI, Amp, Goose, OpenCode, and 30+ others.
# Personal (all your projects) git clone https://github.com/webloomhq/skills.git ~/.claude/skills/webloom # Per-project cd your-project git clone https://github.com/webloomhq/skills.git .claude/skills/webloom # Restart Claude Code — 116 site skills auto-discover. # Say something like "post to LinkedIn" and the LinkedIn skill activates.
The skills bundle NO thread payload (no proven_actions, no selectors). They're triggers + workflow guides that instruct the agent to call install_thread(domain=...) which runs the normal WebLoom flow: trial for 3 days, $5 one-time, or included with WebLoom Pro.
Architecture
Two layers, top to bottom:
JSON knowledge about ONE site. Selectors, endpoints, quirks, click + fill strategies, anti-bot pacing. Merged into the live playbook at runtime; user observations override Thread defaults.
62 tools covering click ladder, 4 upload strategies, React/Redux/AUI/Backbone awareness, vision Layer 2.5, network capture, recipe record/replay. Stays lean — complexity migrates to Threads.
Thread schema
A Thread is a single JSON file with one required field — domain — and a fistful of optional fields the engine knows how to consume:
{
"domain": "kdp.amazon.com",
"name": "KDP Auto-Publish Profile",
"version": "1.0.0",
"author": "username",
"license": "proprietary | mit | cc-by | …",
"framework": "amazon-aui",
"default_strategy": "js",
"notes": [
"AjaxInput clears input.files after onchange — use xhr_upload",
"Modal save needs A.declarative.fire('a:click', target)"
],
"selectors": {
"title_input": "#data-print-book-title",
"categories_button": "#categories-modal-button"
},
"actions": {
"submit_form": {
"type": "aui_dispatch",
"event": "a:click",
"target": "#submit-button"
}
},
"captured_endpoints": [
{ "type": "fetch", "url": "https://kdp.amazon.com/api/…" }
],
"click_log": {
"Choose categories": {
"strategy": "cdp", "successes": 12, "failures": 1, "last_at": 1716000000
}
}
}Sell a thread
As an invited author, submit your .thread.json via webloom_publish_thread() or the website. Curator approves + sets pricing. Authors keep 90% of every sale; WebLoom keeps 10%.
Payments run through Polar (Merchant of Record). They handle global tax, chargebacks, refunds. You handle: nothing operational. Sites drift — buyers report it in one click, and the engine can self-repair a Thread in minutes. Your reputation and new sales stay intact.
Creator dashboard + Polar integration land in v0.2.
Engine tools
62 tools, grouped:
- ›click — actionability + CDP + JS dispatch + vision
- ›fill — React-aware native setter
- ›key_type — 3 modes (keystrokes, insertText, fast)
- ›key_press — Enter, Tab, Esc, Arrows, ...
- ›scroll_tab, screenshot, navigate
- ›upload_file — Strategy A/B/C/D
- ›xhr_upload — direct FormData to endpoint
- ›replay_xhr — replay captured calls
- ›scan_tab — full or ax-tree mode
- ›scan_tab_diff — incremental
- ›framework_detect — React/Redux/AUI/Backbone/Radix
- ›detect_anti_bot — Cloudflare/DataDome/Akamai
- ›detect_blocker — CAPTCHA/2FA prompts
- ›react_force_change — fiber-walk onChange
- ›react_inspect_store — 4-tier discovery
- ›redux_dispatch — direct action dispatch
- ›aui_dispatch — A.declarative.fire
- ›backbone_inspect
- ›touch_tap — CDP Input.dispatchTouchEvent
- ›start_recording / end_recording
- ›replay_recipe
- ›seed_from_tab — author thread from current tab
- ›list_threads / install_thread / export_thread
- ›save_playbook / get_playbook / note
- ›auth_totp — pyotp
- ›pause_for_human — clean handoff
- ›export_profile / import_profile
- ›capture_network_start / stop
- ›get_captured_requests
- ›wait_for / wait_for_idle