A local control center for working with AI agents in real terminals: open projects, assemble teams, assign tasks, follow execution, and review changes in your browser.
The server runs on your machine and uses your files, Git, and AI CLIs. The browser is the interface; agents run on the computer hosting Cockpit.
#Features and limits
| Area | What is available in this checkout |
|---|---|
| Projects | Open folders, prepare repositories, clone, and browse recent projects. |
| Work modes | Office, Command, Code, and Terminals, with Alt+1 through Alt+4; mission Progress with Alt+5. Work modes. |
| Workspace | File explorer, tabs, editor, Git, and project memory, with Orchestration alongside. Code and files. |
| Missions | Create, start, pause, resume, and restore missions with their own objectives and teams. |
| Team | Panels with roles, runners, models, effort, accounts, skills, instructions, connections, and messages. |
| Terminals | xterm interface, real PTY resizing, and a separate session host for recovery after server restarts. |
| Orchestration | Dispatch, task dependencies, panel supervision, autonomy limits, and event tracking. |
| Tasks | States todo, in-progress, blocked, in-review, complete, and failed, file scopes, and run history. |
| Runs | State machine, validated structured results, separate evidence, and disk persistence. |
| Review | Git diffs, files, hunks, annotations, and review decisions. |
| Race | Concurrent attempts at a task, with candidate comparison and winner selection. |
| Isolation | Git worktrees and shared folders, with copy and symlink configuration. |
| Handoff | Sanitized transcripts, ANSI removal, and budget-based truncation for context transfer. |
| Context | Claude and Codex session readings, warnings at 60% and 75%, and New session with summary saved to project memory. |
| Verification | Background typecheck and test after delivery. Failures return to the agent with logs, up to twice. Configure .cockpit/verify.json or use COCKPIT_VERIFY=0. |
| Preview | Automatic setup for web servers with remote access and HMR, static files, or Expo via QR code. Technical settings are under Advanced options. See docs/LIVE_PREVIEW.md. |
| Connectors | MCP connectors with estimated token costs. Imported plugin MCPs and hooks start disabled. Claude Code can also be used as an advanced tool. GET /api/openapi.json describes the API. |
| Library | Roles, profiles, squads, recipes, and skills with persistent activation and automatic discovery. |
| Marketplace | Git sources with plugin manifests, downloads, and skill installation. |
| Activity and topology | Mission, relationship, and task tracking through the interface and WebSocket channels. |
| Compliance | Invariant reports and matrices, plus audit artifact generation. |
| Voice | Browser capture and server-side CPU transcription with Transformers.js/Whisper. |
| Media | Provider detection and configuration; CLI installation when the catalog includes an installer. |
| Remote access | Link/QR invitations, devices, expiry, revocation, and view or control permissions. |
| Companion | Mission summaries using a user-configured Ollama endpoint. |
| DSH | Optional integration with a local DeepSeek Harness checkout and isolated data. |
| Office | Full Habblaud 1.0.2 (MIT), using real Cockpit panels, terminals, and messages, plus a compact accessible list. Office. |
| Live | Two-person call pilot: voice, screen sharing, optional camera, chat, Cockpit sign-in, and floating dock. Live. |
| Personal assistant | Per-project chat with your chosen AI, project memory, Needs attention, and Telegram channels. Assistant and Telegram. |
| Creation guide | Included Vibe Coding Toolkit (MIT), for preparing requests from guides and recipes. |
Current limits: catalog profiles may require runners you have not installed. A listed model does not guarantee your account can access it. Connection discovery relies on local configuration signals; verify authentication in the CLI itself. Media settings configure tools rather than providing a universal generation API. Marketplace MCP metadata does not mean every MCP server is configured automatically for every runner.
Documents in docs/ also describe plans such as Brain, SQLite FTS5, embeddings, and an Obsidian adapter. These are not all shipped features. The engines in this checkout persist data in local files, mainly JSON.