clawdcursor

Clawd Cursor

If a human can do it on a screen, you can too. No API? No integration? No problem. Every app already has a UI — Clawd Cursor gives you eyes and hands to use all of them.

No task is impossible. If it's on the screen, it's within reach.

IMPORTANT — READ THIS BEFORE ANYTHING ELSE

You have a desktop AI agent over GUI running on the user's machine. This means:

1. You CAN open apps, click buttons, type text, fill forms
2. You CAN see what's on screen (via Clawd Cursor's screenshot pipeline)
3. You CAN interact with applications — browser, desktop apps, settings
4. You MUST respect safety tiers — 🟢 Auto actions run freely, 🔴 Confirm actions require user approval
5. You MUST ask before accessing sensitive apps (email, banking, private messages, password managers)
6. You SHOULD prefer direct tools (file read, web_fetch, CLI) over desktop automation when available

Think of Clawd Cursor as your hands and eyes on the user's computer — used with their permission.

---

When to Use This (AGENTS: READ THIS)

You have a desktop UI agent available. Use it when the user asks you to interact with their desktop, or when no other tool can accomplish the task.

Tool vs Skill (OpenClaw terminology)

• Tool = direct capability (API call, filesystem, shell, web fetch, browser command).
• Skill = packaged workflow/domain logic that may call one or more tools.
• This skill (Clawd Cursor) = GUI execution skill. Use it after OpenClaw tools/skills that can complete the same work without GUI.

Use Clawd Cursor for (examples, not limits):

Clawd Cursor can perform any action that is visible and interactable in the GUI (subject to safety policy).

• User-requested desktop tasks — "open Gmail and send an email", "check my calendar"
• Read a webpage — when web_fetch or browser tools aren't available
• Interact with desktop apps — click buttons, fill forms, read results
• Browser tasks — search, navigate, fill forms (when browser tool unavailable)
• Visual verification — did the page load? what does the UI show?
• Cross-app workflows — copy from one app, paste in another
• Settings changes — when the user explicitly asks

⚠️ Sensitive App Policy

Always ask the user before accessing:

• Email clients (Gmail, Outlook)
• Banking or financial apps
• Private messaging (WhatsApp, Signal, Telegram)
• Password managers
• Admin panels or cloud consoles

Don't use Clawd Cursor when:

• You can do it with a direct API call or CLI command (faster)
• The task is purely computational (math, text generation, code writing)
• You can already read/write the file directly
• The browser tool or web_fetch can handle it

OpenClaw + Clawd Cursor Routing Contract (Avoid Overlap)

Clawd Cursor should be treated as OpenClaw's GUI execution layer, not a competing planner.

Route tasks in this order:

1. OpenClaw native tools first (filesystem, API, shell, provider-native skills)
2. Browser-native automation next (Playwright/CDP direct) for browser-only reads/clicks
3. Clawd Cursor API task (POST /task) only when desktop/UI-level interaction is required

Practical rule

• If OpenClaw already has a reliable skill/tool for the domain, use it.
• Use Clawd Cursor to bridge gaps where no API/tool exists or when the user explicitly asks for GUI interaction.

This keeps behavior predictable, lowers latency/cost, and avoids duplicated logic between the main OpenClaw agent and this skill.

Universal task pattern

For broad "get it done" requests, split into three phases:

1. Plan in OpenClaw: break work into API/CLI/browser/GUI subtasks.
2. Execute cheap paths first: API + CLI + browser direct.
3. Escalate only residual UI steps to Clawd Cursor.

Think: "OpenClaw decides, Clawd Cursor acts on GUI when needed."

Direct Browser Access (Fast Path)

For quick page reads without a full task, connect to Chrome via Playwright CDP:

const pw = require('playwright');
const browser = await pw.chromium.connectOverCDP('http://127.0.0.1:9222');
const pages = browser.contexts()[0].pages();
const text = await pages[0].innerText('body');

Use this when you just need page content — faster than sending a task.

---

REST API Reference

Base URL: http://127.0.0.1:3847

Note: On Windows PowerShell, use curl.exe (with .exe) or Invoke-RestMethod. Bare curl is aliased to Invoke-WebRequest which behaves differently.

Pre-flight Check

Before your first task, verify Clawd Cursor is running:

curl.exe -s http://127.0.0.1:3847/health

Expected: {"status":"ok","version":"0.6.0"}

If connection refused — start it yourself (don't ask the user):

# Find the skill directory and start the server
Start-Process -FilePath "node" -ArgumentList "dist/index.js","start" -WorkingDirectory "<clawd-cursor-directory>" -WindowStyle Hidden
Start-Sleep 3
# Verify it's running
curl.exe -s http://127.0.0.1:3847/health

The skill directory is wherever SKILL.md lives (the parent of this file). Use that path as the working directory.

Sending a Task (Async — Returns Immediately)

POST /task accepts the task and returns immediately. The task runs in the background. You must poll /status to know when it's done.

curl.exe -s -X POST http://127.0.0.1:3847/task -H "Content-Type: application/json" -d "{\"task\": \"YOUR_TASK_HERE\"}"

PowerShell:

Invoke-RestMethod -Uri http://127.0.0.1:3847/task -Method POST -ContentType "application/json" -Body '{"task": "YOUR_TASK_HERE"}'

Polling Pattern (Follow This)

1. POST /task → get accepted
2. Wait 2 seconds
3. GET /status
4. If status is "idle" → done
5. If status is "waiting_confirm" → ASK THE USER, then POST /confirm based on their answer
6. If still running → wait 2 more seconds, go to step 3
7. If 60+ seconds → POST /abort and retry with clearer instructions

Checking Status

curl.exe -s http://127.0.0.1:3847/status

Confirming Safety-Gated Actions

Some actions (sending messages, deleting) require approval. 🔴 NEVER self-approve these. Always ask the user for confirmation before POST /confirm. These exist to protect the user — do not bypass them.

curl.exe -s -X POST http://127.0.0.1:3847/confirm -H "Content-Type: application/json" -d "{\"approved\": true}"

Aborting a Task

curl.exe -s -X POST http://127.0.0.1:3847/abort

Reading Logs (Debugging)

curl.exe -s http://127.0.0.1:3847/logs

Returns last 200 log entries. Check for error or warn entries when tasks fail.

Response States

---

CDP Direct Reference

Chrome must be running with --remote-debugging-port=9222.

Quick check:

curl.exe -s http://127.0.0.1:9222/json/version

If this returns JSON, Chrome is ready.

Connecting via Playwright:

const { chromium } = require('playwright');
const browser = await chromium.connectOverCDP('http://127.0.0.1:9222');
const context = browser.contexts()[0];
const page = context.pages()[0];

// Read page content
const title = await page.title();
const url = page.url();
const text = await page.textContent('body');

// Click by role
await page.getByRole('button', { name: 'Submit' }).click();

// Fill a field
await page.getByLabel('Email').fill('user@example.com');

// Read specific elements
const buttons = await page.$eval('button', els => els.map(e => e.textContent));

---

Task Writing Guidelines

1. Be specific — include app names, URLs, exact text to type, button names
2. One task at a time — wait for completion before sending the next
3. Describe the goal, not the clicks — say "Send an email to john@example.com about the meeting" not "click compose, click to field..."
4. Check status if a task seems to hang
5. Don't include credentials in task text — tasks are logged

Task Examples

Error Recovery

---

How It Works — 5-Layer Pipeline

Layer 1 includes keyboard shortcuts — common actions execute as direct keystrokes (0 LLM calls).

80%+ of tasks handled by Layer 0-1 (free, instant). Vision model is last resort only.

Safety Tiers

Security & Privacy

Network Isolation

• API binds to 127.0.0.1 only — not network accessible. Verify: netstat -an | findstr 3847 should show 127.0.0.1:3847
• Screenshots stay in memory, never saved to disk (unless --debug)
• No telemetry, no analytics, no phone-home calls

Data Flow

• With Ollama (local): 100% offline — zero external network calls. No data leaves the machine.
• With cloud providers: screenshots/text are sent to the user's chosen provider API only. No data goes to skill authors, ClawHub, or third parties.
• OpenClaw users: credentials auto-discovered from local config files — no keys stored in skill directory.
• The user controls data flow by choosing their provider. Ollama = fully private.

Agent Autonomy Controls

• 🟢 Auto actions (navigation, reading, opening apps) run without prompting
• 🟡 Preview actions (typing, form filling) are logged before executing
• 🔴 Confirm actions (sending messages, deleting, purchases) always pause for user approval
• Agents must ask the user before accessing sensitive apps (email, banking, messaging, passwords)
• Agents must never self-approve 🔴 Confirm actions

---

Setup (User Reference)

Setup is handled by the user. If Clawd Cursor isn't running, start it yourself using the exec tool:

Start-Process -FilePath "node" -ArgumentList "dist/index.js","start" -WorkingDirectory "<skill-directory>" -WindowStyle Hidden

Only ask the user if you cannot start it (e.g., node not installed, build missing).

git clone https://github.com/AmrDab/clawd-cursor.git
cd clawd-cursor
npm install && npm run build
npx clawd-cursor doctor    # auto-detects and configures everything
npm start                  # starts on port 3847

macOS: Grant Accessibility permission to terminal: System Settings → Privacy & Security → Accessibility

---

Performance Optimization

Proven optimizations applied to reduce task execution latency and LLM API costs. Reference files in perf/references/patches/.

Applied Optimizations

Benchmarks (2560x1440)

Perf Tools

• perf/apply-optimizations.ps1 — apply all patches
• perf/perf-test.ts — benchmark harness (npx ts-node perf/perf-test.ts)