Cursor Agent Skill

Two ways to use Cursor Agent from OpenClaw, for different scenarios.

Related

• OpenClaw Node for VS Code — VS Code/Cursor extension that enables the Node path (install this first for Path 2)
• vscode-node — Standalone skill for the Node path only (if you don't need Cursor CLI)
• Source: github.com/xiaoyaner-home/openclaw-vscode

Path Selection

Scenario	Path	Why
Quick coding task, bug fix, refactor	CLI	Fast, no setup, works anywhere
Generate code, review PR, write tests	CLI	Non-interactive -p mode is perfect
Fix type errors using real diagnostics	Node	diagnostics.get shows actual TS/lint errors
Navigate definitions/references first	Node	lang.definition, lang.references
Run project tests and iterate	Node	test.run + test.results loop
Debug with breakpoints	Node	Full debug protocol
Targeted changes to a specific project	Node	IDE workspace context is precise

Default: CLI. Use Node only when you specifically need IDE intelligence.

---

Path 1: CLI (Local Cursor Agent)

Prerequisites

# Install
curl https://cursor.com/install -fsS | bash

# Login
agent login

# Verify
agent --version

Modes

Mode	Flag	Use Case
Agent	(default)	Full coding — reads, writes, runs commands
Plan	--plan or --mode=plan	Design approach first, then choose local or cloud execution
Ask	--mode=ask	Read-only codebase exploration, no edits

Interactive Mode

# Start interactive session
agent

# Start with prompt
agent "refactor the auth module to use JWT tokens"

# Start in plan mode
agent --plan "design a caching layer for the API"

# Start in ask mode
agent --mode=ask "explain how the auth middleware works"

Non-Interactive Mode (Automation)

# One-shot task (prints result, exits)
agent -p "find and fix all unused imports in src/"

# With specific model
agent -p "review this code for security issues" --model gpt-5.2

# JSON output for parsing
agent -p "list all TODO comments" --output-format json

# Streaming JSON (real-time)
agent -p "run tests and report" --output-format stream-json --stream-partial-output

# Force mode (auto-apply changes, no confirmation)
agent -p "fix all linting errors" --force

Cloud Agent Handoff

Push work to Cursor's cloud to continue running while you're away:

# Start directly in cloud
agent -c "refactor the auth module and add comprehensive tests"

# Mid-conversation: prepend & to send to cloud
& refactor the auth module and add comprehensive tests

Pick up at cursor.com/agents.

Session Management

agent ls              # List previous conversations
agent resume          # Resume most recent
agent --continue      # Continue previous session
agent --resume="id"   # Resume specific conversation

Slash Commands (Interactive)

Command	Action
/plan	Switch to Plan mode / view current plan
/ask	Switch to Ask mode
/models	Switch AI model
/compress	Summarize conversation, free context
/rules	Create/edit rules
/commands	Create/edit custom commands
/mcp enable <name>	Enable MCP server
/mcp disable <name>	Disable MCP server
/sandbox	Configure sandbox mode
/max-mode [on|off]	Toggle Max Mode
/resume	Resume previous conversation

Keyboard Shortcuts

Shortcut	Action
Shift+Tab	Rotate modes (Agent → Plan → Ask)
Shift+Enter	Insert newline (multi-line prompt)
Ctrl+R	Review changes (i for instructions, arrows to navigate)
Ctrl+D	Exit (double-press for safety)
ArrowUp	Cycle previous messages

Context & Rules

The CLI automatically loads: - .cursor/rules directory - AGENTS.md at project root - CLAUDE.md at project root - MCP servers from mcp.json

Use @filename or @directory/ in interactive mode to include context.

⚠️ Using CLI from OpenClaw (PTY Required)

Cursor CLI is an interactive TUI — it needs a real terminal. Use pty:true:

# ✅ Correct — with PTY
exec pty:true command:"agent -p 'Your task'" workdir:/path/to/project

# ✅ Background for longer tasks
exec pty:true background:true command:"agent -p 'Build REST API'" workdir:/path/to/project

# ❌ Wrong — will hang
exec command:"agent -p 'Your task'"

For long tasks, use background + poll:

# Start
exec pty:true background:true workdir:~/project command:"agent -p 'Add comprehensive tests for the auth module' --force"

# Check progress
process action:log sessionId:XXX

# Check if done
process action:poll sessionId:XXX

Sandbox Controls

# Start with sandbox enabled
agent --sandbox enabled

# Start with sandbox disabled
agent --sandbox disabled

# Configure interactively
/sandbox

Sandbox supports granular network access controls — define which domains the agent can reach.

---

Path 2: VS Code / Cursor Node

Remote-control a Cursor/VS Code IDE through the OpenClaw Node protocol. The IDE must have the openclaw-node-vscode extension installed and connected.

Prerequisites

• Extension installed: VS Code Marketplace
• Node visible in nodes status
• Extension status bar shows 🟢

Invocation Pattern

nodes invoke --node "<name>" --invokeCommand "<cmd>" --invokeParamsJson '{"key":"val"}'

Timeout Guide

Operation	invokeTimeoutMs	Notes
File/editor/lang	15000	Fast IDE operations
Git	30000	May involve disk I/O
Test	60000	Depends on test suite
Agent plan/ask	180000	AI thinking time
Agent run	300000	Full coding task

Command Reference

Category	Prefix	Key Commands
File	vscode.file.*	read, write, edit, delete
Directory	vscode.dir.*	list
Language	vscode.lang.*	definition, references, hover, symbols, rename, codeActions, format
Editor	vscode.editor.*	context, openFiles, selections
Diagnostics	vscode.diagnostics.*	get (errors/warnings)
Git	vscode.git.*	status, diff, log, blame, stage, unstage, commit, stash
Test	vscode.test.*	list, run, results
Debug	vscode.debug.*	launch, stop, breakpoint, evaluate, stackTrace, variables, status
Agent	vscode.agent.*	status, run, setup
Workspace	vscode.workspace.*	info

Quick Examples

# Read a file
nodes invoke --node "my-cursor" --invokeCommand "vscode.file.read" 
  --invokeParamsJson '{"path":"src/main.ts"}'

# Get diagnostics (real type errors!)
nodes invoke --node "my-cursor" --invokeCommand "vscode.diagnostics.get"

# Go to definition
nodes invoke --node "my-cursor" --invokeCommand "vscode.lang.definition" 
  --invokeParamsJson '{"path":"src/main.ts","line":10,"character":5}'

# Git status + commit
nodes invoke --node "my-cursor" --invokeCommand "vscode.git.status"
nodes invoke --node "my-cursor" --invokeCommand "vscode.git.stage" 
  --invokeParamsJson '{"paths":["src/main.ts"]}'
nodes invoke --node "my-cursor" --invokeCommand "vscode.git.commit" 
  --invokeParamsJson '{"message":"fix: resolve type error"}'

# Delegate to Cursor Agent (through IDE)
nodes invoke --node "my-cursor" --invokeCommand "vscode.agent.run" 
  --invokeParamsJson '{"prompt":"Add error handling to all API endpoints","mode":"plan"}' 
  --invokeTimeoutMs 180000

Node Workflow: Fix → Verify → Commit

The real power of Node path — a closed loop with IDE intelligence:

1. diagnostics.get           → Find real errors
2. vscode.agent.run (fix)    → Let Cursor Agent fix them
3. diagnostics.get           → Verify errors resolved
4. test.run                  → Run tests
5. test.results              → Check results
6. git.diff                  → Review changes
7. git.stage + git.commit    → Ship it

No tmux, no TTY hacks — all through VS Code API.

---

Combined Workflow Example

Use CLI for broad tasks, switch to Node for precision:

1. CLI: agent -p "implement user authentication module" --force
   → Generates the initial code quickly

2. Node: vscode.diagnostics.get
   → Reveals 3 type errors the CLI missed

3. Node: vscode.agent.run '{"prompt":"fix these type errors: ..."}'
   → Fixes with full IDE context

4. Node: vscode.test.run
   → Runs test suite

5. Node: vscode.git.stage + vscode.git.commit
   → Clean commit

---

Error Handling

Error	Cause	Fix
CLI hangs	No PTY	Add pty:true to exec
node not found	Extension disconnected	Check VS Code status bar
command not allowed	Gateway whitelist	Add to gateway.nodes.allowCommands
timeout	Operation too long	Increase invokeTimeoutMs
path traversal blocked	Absolute path used	Use relative paths for Node

Security

• CLI: Respects sandbox mode, command approval, rules
• Node: All paths relative to workspace, Ed25519 device identity, Gateway approval required
• Both: No raw shell access by default