Introduction

OpenCockpit is the open-source GUI for Claude Code — and a single canvas for whatever AI agent you bring next. Everything runs locally.

What you get

• Multi-project parallel sessions. Run 5+ agent sessions across separate projects at once. Each one lives in its own tab; you get a desktop toast when any finishes.
• Bring any agent. Claude works out of the box. Codex, DeepSeek, Kimi (Moonshot), and local Ollama models are one tab away — paste an API key (or none for Ollama).
• More than chat. A real terminal, Chrome automation, and bubbles for PostgreSQL / MySQL / Redis / Neo4j / Jupyter — all inside one window your agent can drive.
• Code-aware navigation. LSP go-to-definition, a Code Map of function calls, and a /cg slash command that lets the AI explore the project graph via HTTP.

The three panels

Cockpit's whole UI is three panels rendered in parallel. None of them ever unmount — switching is just a CSS translate. Cmd+1 / 2 / 3 flips between them; on touch devices you can swipe.

Panel 1 — Agent (Cmd+1)

The chat surface. Each tab is an independent conversation with whatever engine you picked.

• Engines. Claude / Codex / DeepSeek / Kimi (Moonshot) / Ollama — switch per tab. Claude works out of the box; the rest need a one-time key (or none for Ollama).
• Sessions. Pin a session to the sidebar, fork it to branch the conversation, search across projects via the Session Browser (sidebar icon).
• AI mode slash commands. Type / in the chat input to pick from six modes that rewire how the AI approaches the next reply: /qa (clarify before changing code), /fx (bug evidence chain, analysis only), /ex (deep structured analysis), /go (landing mode — MVP-staged implementation with self-verify), /cg (CodeGraph exploration), /cc (Cockpit-CLI-driven end-to-end verification). Skills installed via SKILL.md show up in the same menu as /skill-name.
• Shell prefix. Any input starting with ! runs as a shell command and its output is appended to the prompt.
• Scheduled tasks. A sidebar panel for one-shot, interval, and cron schedules — run a prompt every morning to summarise yesterday's PRs, every hour to scan release notes, every 5 minutes to babysit a long-running task.
• Skills. SKILL.md files become callable /skill-name commands across all tabs.

Read more about chat →

Panel 2 — Explorer (Cmd+2)

The code surface. Five tabs across the top:

Inside any file:

• Syntax highlighting for every common language.
• Vi mode in the code editor — i, Esc, h/j/k/l, etc.
• Cmd+F to find within the file (regex, case-sensitive, whole word).
• Cmd+P for fuzzy file open across the whole project.
• Right-click on a line → Blame view with author / commit / time per line.
• Cmd+click a symbol → jump to its definition (TypeScript / JavaScript / Python).
• Code Map — visualise function callers and callees as a graph.

Whatever Explorer shows you, the AI can query the same graph via /cg from the Agent panel — see CodeGraph.

Read more about Explorer →

Panel 3 — Console (Cmd+3)

The "everything else" surface. Type into the input bar at the bottom and Cockpit picks the right bubble type for you:

Each bubble is draggable; Cmd+M maximises the focused one. Bubbles persist across sessions until you close them, and the CLI lets external scripts drive them (see CLI Reference).

Read more about Console →

Cross-panel features

• Code references via comments. In Explorer, select a range of code and add a comment (Cmd+/ or the floating toolbar). Add as many as you need across files; when you're ready to ask the AI, the Comments modal in Agent renders every pinned comment as a formatted block (file path + line range + code + your note) that you can copy straight into chat. This is the supported path for "tell the AI about this specific code" — there is no drag-and-drop.
• Jump to file from search. Cmd+P opens fuzzy file search and lands you in the Explorer tree at the matched file, ready to comment / preview / blame.
• Bubbles ↔ Agent. Every Browser bubble (and one-shot / interactive terminal bubble) carries a short-ID badge in its title bar. Click it to register the bubble and copy cock browser <id> / cock terminal <id> to your clipboard — paste it into chat and the AI can drive that exact bubble through Cockpit's CLI.

Who this is for

If you spend your day driving Claude Code through a terminal and wish you could:

• watch five concurrent agent sessions without tmux gymnastics
• swap to Codex or DeepSeek for one task without rebooting your shell
• click a row in the file tree and immediately see Git blame for that line
• have the AI query Postgres and visualise the result without leaving the chat

…then Cockpit is the surface you've been hand-rolling in scripts.

What this is not

• Not a hosted product. Everything runs locally; you bring your own keys. There is no SaaS plan.
• Not a code editor. Cockpit views and reviews code, but a deep editing experience belongs in VS Code / Cursor / nvim. We integrate cleanly with whatever editor you already use.
• Not a Claude wrapper. Claude is the default, but every panel is engine-agnostic. The codebase already ships five model adapters.

Next steps

• Quickstart — install, run, and a realistic end-to-end task using all three panels
• Skills — the single page most users wish they'd read first
• Engines Overview — add Codex / DeepSeek / Kimi / Ollama as alternative tabs