cursor-cloud-agents（Cursor cloud Agent 搜索）

Cursor Cloud Agents Skill

⚡ Quick Reference

Most common commands and patterns:

# Launch an agent (uses default model: gpt-5.2)
cursor-api.sh launch --repo owner/repo --prompt "Add tests for auth module"

# Check agent status
cursor-api.sh status <agent-id>

# Get conversation history
cursor-api.sh conversation <agent-id>

# Send follow-up message
cursor-api.sh followup <agent-id> --prompt "Also add edge case tests"

# List all agents
cursor-api.sh list

# Check usage/quota
cursor-api.sh usage

Common Options:

• --model <name> - Specify model (default: gpt-5.2)
• --branch <name> - Target branch
• --no-pr - Don't auto-create PR
• --no-cache - Bypass cache
• --verbose - Debug output
• --background - Run agent in background mode

Background Tasks:

cursor-api.sh launch --repo owner/repo --prompt "..." --background
cursor-api.sh bg-list
cursor-api.sh bg-status <task-id>
cursor-api.sh bg-logs <task-id>

Max Runtime (Background Tasks):

# Default is 24 hours
 cursor-api.sh launch --repo owner/repo --prompt "..." --background

# Custom max runtime (2 hours)
cursor-api.sh launch --repo owner/repo --prompt "..." --background --max-runtime 7200

# Unlimited runtime (not recommended)
cursor-api.sh launch --repo owner/repo --prompt "..." --background --max-runtime 0

# Set default via environment variable
export CURSOR_BG_MAX_RUNTIME=43200  # 12 hours
cursor-api.sh launch --repo owner/repo --prompt "..." --background

Short Commands (cca aliases):

For faster daily usage, source the cca-aliases.sh file:

source scripts/cca-aliases.sh

Then use cca instead of cursor-api.sh:

cca list                    # List agents
cca launch --repo ...       # Launch agent
cca status <id>             # Check status
cca conversation <id>       # Get conversation
cca followup <id> --prompt  # Send followup
cca delete <id>             # Delete agent

Exit Codes: 0=Success, 1=API Error, 2=Auth, 3=Rate Limit, 4=Repo Access, 5=Invalid Args

Overview

This skill wraps the Cursor Cloud Agents HTTP API, allowing OpenClaw to dispatch coding tasks to Cursor's cloud agents, monitor their progress, and incorporate results.

When to Use

Use this skill when you need to:

• Delegate coding tasks to Cursor agents running on GitHub repositories
• Generate code, tests, or documentation on existing codebases
• Perform refactoring or feature implementation asynchronously
• Get a "second opinion" on code changes

When NOT to Use

• For simple questions that don't require code changes
• When you need real-time streaming responses (use local Cursor CLI instead)
• For tasks outside of GitHub repositories

Authentication

The skill automatically discovers your Cursor API key from these locations (in order):

1. Environment variable: CURSOR_API_KEY
2. OpenClaw env file: ~/.openclaw/.env
3. OpenClaw local env: ~/.openclaw/.env.local
4. Project env: .env in current directory
5. Cursor config: ~/.cursor/config.json

Recommended: Add to ~/.openclaw/.env:

CURSOR_API_KEY=your_cursor_api_key_here

To get your API key:

1. Open Cursor IDE
2. Go to Settings → General
3. Copy your API key

Verify it's working:

cursor-api.sh me

Workflow Patterns

Pattern A: Fire-and-Forget

Launch an agent and let it work independently. Check back later.

# Launch agent (uses default model: gpt-5.2)
cursor-api.sh launch --repo owner/repo --prompt "Add comprehensive tests for auth module"

# Launch with specific model
cursor-api.sh launch --repo owner/repo --prompt "Add tests" --model claude-4-opus

# Response: {"id": "agent_123", "status": "CREATING", ...}

# Later - check status
cursor-api.sh status agent_123

Note: If no --model is specified, the default model (gpt-5.2) will be used automatically. You'll see a message indicating which model is being used.

Best for: Tasks that don't need immediate attention, exploratory work

Pattern B: Supervised Dispatch

Launch, monitor, and report results when complete.

# 1. Launch
cursor-api.sh launch --repo owner/repo --prompt "Implement user authentication"

# 2. Poll for completion (check every 60 seconds)
while true; do
    status=$(cursor-api.sh status agent_123)
    if [[ $(echo "$status" | jq -r '.status') == "FINISHED" ]]; then
        break
    fi
    sleep 60
done

# 3. Get results
cursor-api.sh conversation agent_123 | jq -r '.messages[] | select(.role == "assistant") | .content'

Best for: Important tasks where you want to report completion

Pattern C: Iterative Collaboration

Launch, review, and send follow-ups to refine work.

# 1. Launch initial task
cursor-api.sh launch --repo owner/repo --prompt "Add login page"

# 2. Review conversation
cursor-api.sh conversation agent_123

# 3. Send follow-up
cursor-api.sh followup agent_123 --prompt "Also add form validation and error handling"

# 4. Final review when done
cursor-api.sh conversation agent_123

Best for: Complex tasks requiring multiple iterations

Pattern D: Background Mode

For long-running tasks, launch agents in background mode and check on them later.

# Launch in background
result=$(cursor-api.sh launch --repo owner/repo --prompt "Refactor entire codebase" --background)
task_id=$(echo "$result" | jq -r '.background_task_id')
echo "Task started: $task_id"

# List active background tasks
cursor-api.sh bg-list

# Check specific task status
cursor-api.sh bg-status $task_id

# View logs
cursor-api.sh bg-logs $task_id

# List all tasks including completed ones
cursor-api.sh bg-list --all

Background tasks are monitored automatically and logs are saved to ~/.cache/cursor-api/background-tasks/.

Best for: Long-running tasks (10+ minutes), batch operations, CI/CD integration

Commands Reference

List Agents

cursor-api.sh list

Returns all agents with status, repo, and creation time.

Launch Agent

cursor-api.sh launch --repo owner/repo --prompt "Your task description" [--model model-name] [--branch branch-name] [--no-pr] [--background]

Options:

• --repo (required): Repository in owner/repo format
• --prompt (required): Initial instructions for the agent
• --model (optional): Model to use (defaults to gpt-5.2 if not specified)
• --branch (optional): Target branch name (auto-generated if omitted)
• --no-pr (optional): Don't auto-create a PR
• --background (optional): Run agent in background mode

Note: When launched without --model, the skill automatically uses gpt-5.2 and displays a message indicating which model is being used.

Background Mode: When using --background, the command returns immediately with a background_task_id. Use bg-list, bg-status, and bg-logs to monitor progress.

Check Status

cursor-api.sh status <agent-id>

Returns:

• status: CREATING, RUNNING, FINISHED, STOPPED, ERROR
• summary: Summary of work done (if finished)
• prUrl: URL to created PR (if any)

Get Conversation

cursor-api.sh conversation <agent-id>

Returns full message history including all prompts and responses.

Send Follow-up

cursor-api.sh followup <agent-id> --prompt "Additional instructions"

Resumes a stopped or finished agent with new instructions.

Stop Agent

cursor-api.sh stop <agent-id>

Stops a running agent gracefully.

Delete Agent

cursor-api.sh delete <agent-id>

Permanently deletes an agent and its conversation history.

List Models

cursor-api.sh models

Returns available models for agent tasks.

Account Info

cursor-api.sh me

Returns account information including subscription tier.

Verify Repository

cursor-api.sh verify owner/repo

Checks if the specified repository is accessible by Cursor agents.

Exit code 4 if repository not accessible.

Usage/Cost Tracking

cursor-api.sh usage

Returns usage information including:

• Agents used vs. limit
• Compute consumption
• Subscription tier

Clear Cache

cursor-api.sh clear-cache

Clears the response cache.

Background Task Commands

cursor-api.sh bg-list [--all]

List background tasks. By default, excludes completed tasks. Use --all to include finished tasks.

cursor-api.sh bg-status <task-id>

Get detailed status of a background task including current agent state.

cursor-api.sh bg-logs <task-id>

Show logs for a background task. Logs include status changes and any PR URLs created.

Rate Limiting

The skill enforces a 1 request per second rate limit locally to avoid API rate limits. This is applied automatically to all API calls.

If you hit Cursor's API rate limit (HTTP 429), the script exits with code 3.

Caching

GET requests (list, status, conversation, models, me) are cached for 60 seconds by default. To disable caching for a command:

cursor-api.sh --no-cache status agent_123

To change the cache TTL, set the environment variable:

export CURSOR_CACHE_TTL=120  # 2 minutes
cursor-api.sh status agent_123

Exit Codes

Testing

The skill includes a comprehensive test suite (cca-comprehensive-test.sh) that validates:

• Authentication: Auto-discovery, missing key, invalid key handling
• Account Commands: me, usage, models
• Agent Lifecycle: list, launch (with/without model), status, conversation, followup, stop
• Error Handling: Invalid formats, missing args, non-existent agents (all return correct exit codes)
• Options: --verbose, --no-cache, pagination

All tests pass with proper exit codes. Error conditions are correctly handled and return appropriate exit codes.

Concurrent Agent Limits

Based on available documentation and API behavior, Cursor Cloud Agents have the following limits:

These limits are enforced at the account level across all agents. If you exceed the limit, the API returns HTTP 429 with code CONCURRENT_LIMIT.

To check your current usage:

cursor-api.sh usage | jq '.usage.agentsUsed, .limits.concurrentAgents'

Best practices:

1. Stop finished agents when no longer needed
2. Use cursor-api.sh list to monitor active agents
3. Consider batching work into fewer, larger agents rather than many small ones

Note: Concurrent limits are subject to change. Check cursor-api.sh usage for your current account limits.

Best Practices

1. Always Verify Repository Access

Before launching, verify the repository is accessible:

if cursor-api.sh verify owner/repo >/dev/null 2>&1; then
    cursor-api.sh launch --repo owner/repo --prompt "..."
else
    echo "Repository not accessible. Install the Cursor GitHub App."
fi

2. Use Clear, Specific Prompts

Good prompt:

"Add comprehensive unit tests for the auth module in src/auth/, covering login, logout, and token refresh. Use Jest and mock external API calls."

Bad prompt:

"Add some tests"

3. Check Usage Before Launching

Monitor your quota:

cursor-api.sh usage | jq '.usage'

4. Clean Up Finished Agents

Delete agents you no longer need:

cursor-api.sh list | jq -r '.[] | select(.status == "FINISHED") | .id' | while read id; do
    cursor-api.sh delete "$id"
done

5. Choose Appropriate Max Runtime for Background Tasks

Typical task durations:

• Quick fixes (typos, small bugs): 5-15 minutes → --max-runtime 900
• Feature implementation: 30-60 minutes → --max-runtime 3600
• Large refactors: 2-6 hours → --max-runtime 21600
• Complex migrations: 6-24 hours → --max-runtime 86400 (default)

# Check remaining time
 cursor-api.sh bg-status <task-id> | jq '.remaining_seconds'

# Set custom max runtime
 cursor-api.sh launch --repo owner/repo --prompt "Migrate database schema" --background --max-runtime 43200  # 12 hours

6. Handle Errors Gracefully

Always check exit codes in scripts:

if ! response=$(cursor-api.sh launch --repo owner/repo --prompt "..." 2>&1); then
    case $? in
        2) echo "Authentication error - check CURSOR_API_KEY" ;;
        3) echo "Rate limited - try again later" ;;
        4) echo "Repository not accessible" ;;
        *) echo "API error: $response" ;;
    esac
fi

Follow-up Templates

Use these templates for common follow-up scenarios:

"Add more tests"

Also add tests for edge cases: empty input, null values, and maximum length limits.

"Fix the implementation"

The current implementation doesn't handle [specific case]. Please update it to [requirement].

"Add documentation"

Add comprehensive JSDoc comments to all public functions and a brief README section explaining the feature.

"Refactor for clarity"

Refactor the code to use more descriptive variable names and extract complex logic into helper functions.

Companion Setup: CLI Backend

For local tasks (not on GitHub repos), also configure the Cursor Agent CLI as a cliBackend:

// In your OpenClaw config
{
  "cliBackends": {
    "cursor-agent": {
      "command": "agent",
      "args": ["-p", "--force", "--output-format", "text"],
      "output": "text",
      "input": "arg",
      "env": {
        "CURSOR_API_KEY": "${CURSOR_API_KEY}"
      }
    }
  }
}

This enables cursor-agent as a backend for local file operations, while this skill handles Cloud Agents for GitHub repos.

Troubleshooting

"Repository not accessible" (exit code 4)

1. Ensure the Cursor GitHub App is installed on the repository
2. Check that you have admin/write access to the repo
3. Verify the repo name is correct (owner/repo format)

"Authentication failed" (exit code 2)

1. Check that CURSOR_API_KEY is set in your environment
2. Verify the API key is valid in Cursor IDE settings
3. Ensure the key hasn't expired

"Rate limited" (exit code 3)

1. Wait a few seconds and retry
2. Check your usage with cursor-api.sh usage
3. Consider stopping unused agents

Agent stuck in "CREATING" status

Agents may take 1-2 minutes to start. If stuck longer:

1. Check Cursor status page for outages
2. Try stopping and relaunching
3. Contact Cursor support if persistent