what is CADE?
Overview
CADE is an agent-first IDE. Not a code editor with AI bolted on. The AI agent takes centre stage, and the entire interface is built around improving the integration between you and the agent.
In a traditional IDE, you write code and occasionally ask AI for help. In CADE, the agent works and you supervise. The file tree, viewer, and terminal exist to give you and the agent shared context.
why build a new IDE for this?
Design Philosophy
Read docs/technical/design/visual-design-philosophy.md
Existing tools treat AI as an assistant inside a human-first editor. That's backwards when the agent is doing 90% of the work. CADE flips the model:
• The terminal is the primary workspace, not a secondary panel
• The file tree gives both you and the agent a shared view of the project structure
• The viewer renders plans, documentation, and code so you can review what the agent is doing
• Everything is keyboard-driven, because power users shouldn't reach for a mouse
The design philosophy is terminal-rooted, not terminal-bound. It borrows from tmux and vim because those tools were built for people who live in their terminal. CADE extends that into the agent era.
what does a typical workflow look like?
Workflow
You open a project tab. The agent starts working in the terminal pane. As it reads and writes files, the file tree updates in real time so you can see what's being touched.
The viewer pane auto-displays plan files via hooks. When the agent creates or updates a plan, it appears in the viewer automatically. You read the plan while the agent executes it. No switching windows, no context loss.
File Tree
watch files
being edited
Agent Terminal
agent works here
you supervise
Plan Viewer
review plans
read docs
Multi-project tabs mean you can run agents across several codebases simultaneously. Sessions persist across browser refreshes and disconnections, so you never lose work.
CADE is designed to pair with nkrdn (a code knowledge graph). nkrdn indexes the codebase structure, and the agent queries it for context. Together they give the agent deep understanding of any project it's working in.
what are the key features?
Features
Hook system — configurable triggers that respond to agent events. Plans auto-display in the viewer. File changes highlight in the tree. Custom hooks can run scripts, update dashboards, or notify you.
Dual terminal — toggle between the agent (Claude Code) and a regular shell with ctrl+a s. Run your own commands without interrupting the agent's session.
Enhanced chat pane — a REPL-style conversation view with sticky input, tool call display, and markdown rendering. See exactly what the agent is thinking and doing.
Remote connections — SSH tunnels with token auth. Run CADE on a beefy server and connect from a laptop. The agent gets server resources, you get a lightweight client.
Session persistence — terminal scrollback, file tree state, viewer position, layout proportions, all saved. Close and reopen, everything is exactly where you left it.
Cross-platform — runs in any browser, plus native desktop apps for Windows, macOS, and Linux via Tauri v2. Mobile support with touch toolbar and slide-out panels.
what about the modes?
Modes & Orchestration
Four agent modes control what Claude can do. Switch with slash commands or cycle with ctrl+a . / ,
Code /code — full read-write access. The agent makes changes to files.
Architect /plan — read-only analysis. The agent plans but can't modify anything.
Review /review — read-only code review. Focused on evaluating existing work.
Orchestrator /orch — coordinate multiple worker agents from a single session.
In Orchestrator mode, you direct a coordinator agent that spawns and manages worker agents. Each worker gets its own task and runs independently. The right pane switches to an agent overview showing live status cards for every worker:
• Agent label, task description, and state indicator
• Live preview of each agent's output
• Approval/rejection controls when agents request review
• Click any card to switch the centre pane to that agent
• Cycle between agents with ctrl+a [ / ]
Agent states: pending → starting → idle → busy → review → done
how is it built?
Stack
Grep package.json — dependencies
frontend — TypeScript, Vite, xterm.js with WebGL rendering
backend — Python 3.11+, FastAPI, uvicorn, PTY management
desktop — Tauri v2 (Rust), cross-platform native binaries
markdown — mertex.md (highlight.js + KaTeX + Mermaid)
The keyboard system uses a tmux-style prefix (ctrl+a) with a 2-second window. Vim keys everywhere: file tree, viewer scrolling, theme selector. Configurable via keybindings.toml.