- Run up to 25 AI agents in parallel: Each task runs in isolation within its own Git worktree (`capy/<task-slug>-<id>` branch) in an independent, sandboxed cloud VM. Process an entire sprint concurrently without merge conflicts or context interference between agents. - Captain + Build + Review triple-agent architecture: Captain is a read-only planning agent that analyzes the codebase and generates detailed specifications (cannot edit files or execute commands). Build is an execution agent that performs file editing, command execution, package installation, and commits. Review automatically performs AI-based code review by adding inline comments to PRs. - Multi-model support: Freely switch between major frontier models such as Claude Opus 4.6/Sonnet 4.5, GPT-5.2/5.3 Codex, Grok 4.1 Fast, Gemini 3 Pro, GLM 4.7, Kimi K2.5, and Qwen 3 Coder on a task-by-task basis. Run the same task with multiple models in parallel to compare results. - GitHub / Slack / Linear / Vercel integration: Automatically generate and manage PRs by connecting to GitHub, create tasks directly from Slack messages, delegate Linear issues, and display them in the Vercel Preview dashboard. - Skills and Agent Instructions configuration: Customize agent coding standards and review rules on a per-project basis using the `.capy/settings.json` configuration file and reusable skill packages. - Context compaction: When approaching token limits in long-running tasks, automatically compress the context while preserving key facts, constraints, and the next action to maintain task continuity. - VM isolation security: Each task runs in an isolated Ubuntu 22.04 VM, and the VM is destroyed after the task is completed. OAuth tokens are stored in an encrypted manner, and only authorized repositories are accessed. Enterprise compliance is available upon request.