At a glance
An open-source agent for IDEs and the CLI with broad provider and local-model support, MCP, browser use, checkpoints, read-only research subagents, persistent agent teams, lifecycle plugins, and configurable approval workflows.
Good choice if
- IDE users who want to approve actions and inspect diffs
- Teams mixing interactive, headless, and multi-agent workflows
- Provider-flexible setups with checkpoints, browser use, and MCP
Check before choosing
- The normal IDE and CLI runtimes execute on the host without a documented OS or container sandbox; Kanban worktrees isolate file trees, not processes, credentials, or network access
- IDE use begins approval-oriented, but the CLI prompt path defaults to auto-approve=true. Version 4.1.8 removes the cosmetic YOLO toggle and makes auto-approve the sole unattended-control surface, so unattended use needs an explicit action policy
- Agent teams, the SDK, and Kanban are powerful but fast-moving; teams do not yet apply to the VS Code or JetBrains extensions and Kanban is a research preview
See 5 more considerations
- The 4.1.x stable VSIX contains legacy and SDK-based extension bundles behind a staged remote rollout; assignments apply on reload, and tasks created on the new extension are hidden while the legacy bundle is active
- Read-only subagents are experimental and intentionally cannot edit files or run arbitrary write workflows
- Shadow-Git checkpoints restore files and conversation state, including files that were untracked when the snapshot was taken, but do not reverse external side effects; large repositories can incur storage and snapshot overhead
- SDK, CLI, and Kanban plugins can install executable JavaScript or TypeScript plus npm or Git dependencies; they require source review and do not currently extend the VS Code or JetBrains products
- OpenTelemetry exports logs and metrics rather than proving task correctness, and the checked-in evaluation framework has disabled or unfinished CI layers with no immutable result record