At a glance
Factory's proprietary agent combines an interactive CLI, structured headless execution, IDE integrations, many hosted or local models, MCP, custom subagents, worktrees, and tiered permission levels from read-only analysis to deployment workflows.
Good choice if
- Composable terminal and CI workflows that need structured streams and explicit autonomy tiers
- Parallel changes isolated through named Git worktrees and custom delegated workers
- Teams combining Factory-hosted models, BYOK providers, local Ollama, MCP, and IDE context
Check before choosing
- The Beta OS sandbox is opt-in; its default per-command mode confines shell children while the main Droid process remains outside the OS boundary
- Sandbox file reads allow all paths unless explicitly denied, allowed network domains remain exfiltration channels, and Factory recommends a container or VM for untrusted code
- Git worktrees isolate file changes, not processes; dirty headless worktrees are preserved for review rather than silently removed
See 5 more considerations
- Droid Control provides browser and desktop automation through a plugin rather than the minimal core tool set
- Hooks run automatically with the current environment's credentials and may block tools or merge in plugin-provided code, so project and plugin hooks must be reviewed as trusted execution
- Session rewind and file snapshots do not reverse external side effects, and persistent knowledge relies on explicit AGENTS.md or memory files rather than learned product memory
- CLI session mirroring to Factory web is enabled by default and must be disabled with cloudSessionSync when local-only conversation storage is required
- Factory publishes selected benchmark scorecards, but its public repository does not expose the closed agent implementation or a complete immutable evaluation corpus