Cursor Cloud Agents Skill

⚡ Quick Reference

Most common commands and patterns:

CODEBLOCK0

Common Options:

• - --model <name> - Specify model (default: gpt-5.2)
• INLINECODE1 - Target branch
• INLINECODE2 - Don't auto-create PR
• INLINECODE3 - Bypass cache
• INLINECODE4 - Debug output
• INLINECODE5 - Run agent in background mode

Background Tasks:
CODEBLOCK1

Max Runtime (Background Tasks):
CODEBLOCK2

Short Commands (cca aliases):

For faster daily usage, source the cca-aliases.sh file:
CODEBLOCK3

Then use cca instead of cursor-api.sh:
CODEBLOCK4

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. 1. Environment variable: INLINECODE8
2. OpenClaw env file: INLINECODE9
3. OpenClaw local env: INLINECODE10
4. Project env: .env in current directory
5. Cursor config: INLINECODE12

Recommended: Add to ~/.openclaw/.env:
CODEBLOCK5

To get your API key:

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

Verify it's working:
CODEBLOCK6

Workflow Patterns

Pattern A: Fire-and-Forget

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

CODEBLOCK7

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.

CODEBLOCK8

Best for: Important tasks where you want to report completion

Pattern C: Iterative Collaboration

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

CODEBLOCK9

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.

CODEBLOCK10

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

CODEBLOCK11

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

Launch Agent

CODEBLOCK12

Options:

• - --repo (required): Repository in owner/repo format
• INLINECODE19 (required): Initial instructions for the agent
• INLINECODE20 (optional): Model to use (defaults to gpt-5.2 if not specified)
• INLINECODE22 (optional): Target branch name (auto-generated if omitted)
• INLINECODE23 (optional): Don't auto-create a PR
• INLINECODE24 (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

CODEBLOCK13

Returns:

• - status: CREATING, RUNNING, FINISHED, STOPPED, ERROR
• INLINECODE33: Summary of work done (if finished)
• INLINECODE34: URL to created PR (if any)

Get Conversation

CODEBLOCK14

Returns full message history including all prompts and responses.

Send Follow-up

CODEBLOCK15

Resumes a stopped or finished agent with new instructions.

Stop Agent

CODEBLOCK16

Stops a running agent gracefully.

Delete Agent

CODEBLOCK17

Permanently deletes an agent and its conversation history.

List Models

CODEBLOCK18

Returns available models for agent tasks.

Account Info

CODEBLOCK19

Returns account information including subscription tier.

Verify Repository

CODEBLOCK20

Checks if the specified repository is accessible by Cursor agents.

Exit code 4 if repository not accessible.

Usage/Cost Tracking

CODEBLOCK21

Returns usage information including:

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

Clear Cache

CODEBLOCK22

Clears the response cache.

Background Task Commands

CODEBLOCK23
List background tasks. By default, excludes completed tasks. Use --all to include finished tasks.

CODEBLOCK24
Get detailed status of a background task including current agent state.

CODEBLOCK25
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:

CODEBLOCK26

To change the cache TTL, set the environment variable:

CODEBLOCK27

Exit Codes

Code	Meaning
0	Success
1

API error (including non-existent resources) | | 2 | Authentication missing or invalid | | 3 | Rate limited | | 4 | Repository not accessible | | 5 | Invalid arguments |

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:

Tier	Concurrent Agents	Notes
Free	1	Limited to basic models
Pro

3 | Access to most models |
| Ultra | 5 | Full model access, priority queue |

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:
CODEBLOCK28

Best practices:

1. 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:

CODEBLOCK29

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:

CODEBLOCK30

4. Clean Up Finished Agents

Delete agents you no longer need:

CODEBLOCK31

5. Choose Appropriate Max Runtime for Background Tasks

Typical task durations:

• - Quick fixes (typos, small bugs): 5-15 minutes → INLINECODE45
• Feature implementation: 30-60 minutes → INLINECODE46
• Large refactors: 2-6 hours → INLINECODE47
• Complex migrations: 6-24 hours → --max-runtime 86400 (default)

CODEBLOCK32

6. Handle Errors Gracefully

Always check exit codes in scripts:

CODEBLOCK33

Follow-up Templates

Use these templates for common follow-up scenarios:

"Add more tests"

CODEBLOCK34

"Fix the implementation"

CODEBLOCK35

"Add documentation"

CODEBLOCK36

"Refactor for clarity"

CODEBLOCK37

Companion Setup: CLI Backend

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

CODEBLOCK38

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. 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. 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. 1. Wait a few seconds and retry
2. Check your usage with INLINECODE53
3. Consider stopping unused agents

Agent stuck in "CREATING" status

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

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