AI agents that work on GitHub repos. They read code, make changes, create branches, and open PRs. You must review and merge the PRs yourself.
export CURSOR_API_KEY=your_api_key # Get from https://cursor.com/settings
cloud-agent launch --plan plan.md
# Output: https://cursor.com/agents?id=bc_abc123cloud-agent launch --plan - <<'EOF'
fix: update error message
- Change text in src/auth/login.ts
EOFAuto-selected based on plan content. Override with --model:
composer-1(fast) - Simple tasks: bug fixes, typosclaude-4.5-opus-high(smart) - Complex tasks: refactors, architecture changes
Only launch tasks that modify completely different files. If two plans touch the same file (even different parts), run them sequentially.
Parallel OK:
- Plans that modify completely different files
- Plans that create new files without modifying existing files
NOT OK:
- Multiple plans modifying the same file (even different functions)
- Plans that depend on outputs from other plans
Use a phased approach:
# Phase 1
cloud-agent launch --plan phase1.md
cloud-agent watch $AGENT_ID
# Phase 2 (after Phase 1 completes)
cloud-agent launch --plan phase2.mdBlock until agent completes:
cloud-agent watch $AGENT_ID --verbose
cloud-agent watch $ID1 $ID2 # Multiple agentsSend instructions to running agents:
cloud-agent followup <agent-id> --messages "Add tests"
cloud-agent followup <agent-id> --messages @instructions.mdcloud-agent conversation <agent-id> # View conversation
cloud-agent open <agent-id> # Open in browser
cloud-agent open <agent-id> --pr # Open PR
cloud-agent delete <agent-id>
cloud-agent cancel <agent-id>
cloud-agent batch-delete --status FINISHED --force