Features
Parallel task execution in isolated cloud sandboxes
Powered by codex-1 (OpenAI o3 optimized for SWE)
Reinforcement learning for clean, human-like patches
Real-time progress monitoring with citations and logs
Iterative test execution until passing
AGENTS.md file support for project guidance
Automatic code commit after task completion
User-configurable environment matching dev setup
GitHub pull request proposal
Read and edit files, run commands (test, lint, typecheck)
Optional internet access (enabled June 2025)
Secure isolated containers (internet disabled by default)
Task completion in 1-30 minutes
Verifiable evidence of actions (citations, terminal logs, test outputs)
192k token context length at medium reasoning effort
Agent Command Center for local & cloud agents
Supercomplete predicts next thought, not just edit
Fast Context for millisecond codebase retrieval
Built-in security review for every pull request
Agent Client Protocol (ACP) for cross-model agents
Spaces for sharing context and Git worktrees across agents
Free SWE-1.6 coding model (limited speed on free tier)
Devin Local plugin system (preview) for subagent MCP tools
Terminal allow/deny lists via CLI scopes
Agent/Editor mode toggle with Cmd+. shortcut
Supports Claude Agent, Codex, OpenCode, Cascade
Kimi K2.7 and GLM 5.2 models added (free until Jul 5)
Claude Sonnet 5 available (30% less quota until Aug 31)
Claude Fable 5 available (re-added July 2026)
ACU usage display (v3.3.18)