cabinet

️ Cabinet

Your knowledge base. Your AI team.
️ Files on disk  •  AI workspaces  •  Agents with memory

The AI-first startup OS where everything lives as markdown files on disk. No database. No vendor lock-in. Self-hosted. Your data never leaves your machine.

Built by Hila Shmuel, former Engineering Manager at Apple — now building Cabinet in public, with the open-source community.

@HilaShmuel  •   runcabinet.com  •   [email protected]

From zero to AI team in 2 minutes

npx create-cabinet@latest
cd cabinet
npm run dev:all

Open http://localhost:4000. The onboarding wizard builds your custom AI team in 5 questions.

Install, update, uninstall

Cabinet runs entirely through npx — no global install needed. The CLI is the cabinetai package; create-cabinet is a thin wrapper around it.

Install / create

npx create-cabinet@latest          # create a cabinet and start it
npx cabinetai create my-startup    # just create, don't start
npx cabinetai run                  # start Cabinet in the current dir

On first run, Cabinet downloads the prebuilt app bundle to ~/.cabinet/app/v{version}/. Your cabinet directory is just a folder of markdown files — put it anywhere.

Update

npx cabinetai update               # check for and install a newer app version

The CLI compares your installed app version against cabinet-release.json from the latest GitHub Release.

Uninstall / remove

npx cabinetai uninstall            # remove cached app versions only
npx cabinetai uninstall --all      # also remove global state + telemetry data
npx cabinetai uninstall --yes      # skip the confirmation prompt
npx cabinetai remove               # alias for uninstall

The command prints a summary of what will be deleted and asks for confirmation before doing anything. Your cabinet directories and their data are never touched — those you'd delete manually.

--all additionally removes the platform-specific telemetry directory:

• macOS: ~/Library/Application Support/cabinet-telemetry/
• Windows: %APPDATA%\cabinet-telemetry\
• Linux: $XDG_CONFIG_HOME/cabinet/ (falls back to ~/.config/cabinet/)

To wipe Cabinet completely, run uninstall --all and then rm -rf your cabinet directories yourself.

See docs/CABINETAI.md for the full CLI reference.

The problem

Every time you start a new Claude session, it forgets everything. Your project context, your decisions, your research — gone. Scattered docs in Notion. AI sessions with no memory. Manual copy-paste between tools.

The solution

One knowledge base. AI agents that remember everything. Scheduled jobs that compound. Your team grows while you sleep.

If it feels like enterprise workflow software, it's wrong. If it feels like watching a team work, it's right.

Philosophy

Cabinet is built around a few principles that we think matter deeply for the future of AI + data tools:

• Yours — Your data stays yours: local, visible, and portable. It’s not trapped inside a particular AI provider’s system with no clean way to get it out. You stay in control of your information.
• Git everything — Memory should have history. You should be able to inspect changes, revert mistakes, audit how knowledge evolves, and treat your AI system like the important infrastructure it is.
• BYOAI — Bring your own AI. Cabinet should work with Claude, Codex, OpenCode, local models, and whatever comes next, without forcing your knowledge into a single provider’s ecosystem.
• KISS — Keep it simple, stupid. AI tools should be understandable, inspectable, and hackable. We prefer plain files, clear behavior, and systems that developers can actually reason about.
• Security — We care deeply about security. If AI is going to work with your documents, research, plans, and internal context, the system should minimize surprise, reduce unnecessary exposure, and make trust a design requirement rather than an afterthought.
• Self-hosted — If AI is going to hold your context, plans, research, and operating memory, it should run in an environment you control.

Everything you need. Nothing you don't.

Ship HTML apps inside your knowledge base

This is the biggest difference between Cabinet and tools like Obsidian or Notion. Drop an index.html in any directory — it renders as an embedded app. Full-screen mode with sidebar auto-collapse. AI-generated apps written directly into your KB. Version controlled via git. No build step.

Not another note-taking app

Hire your AI team in 5 questions

Cabinet ships with 20 pre-built agent templates. Each has a role, recurring jobs, recommended skills, and a workspace in the knowledge base.

How it works

1. Install & Run — One command. Next.js + daemon start.
2. Answer 5 Questions — Cabinet builds your custom AI team.
3. Watch Your Team Work — Agents create missions, write content, scout Reddit, file reports.
4. Knowledge Compounds — Every agent run, every edit adds to the KB. Context builds over time.

AI Runtime Today

Cabinet no longer treats the browser terminal as the only way to run AI work.

• Tasks, jobs, and heartbeats now run through a provider adapter layer with persisted conversations and transcript-driven live views.
• Per-run overrides can choose provider, model, and reasoning effort, while personas and jobs can still inherit defaults.
• Current defaults are structured local adapters: claude_local for Claude Code and codex_local for Codex CLI.
• The web terminal is staying as a first-class interactive surface for direct CLI sessions and future terminal-native features such as Cabinet-managed tmux-like workspaces.

Architecture

cabinet/
  src/
    app/api/         -> Next.js API routes
    components/      -> React components (sidebar, editor, agents, jobs, terminal)
    stores/          -> Zustand state management
    lib/             -> Storage, markdown, git, agents, jobs
  server/
    cabinet-daemon.ts -> WebSocket + job scheduler + structured adapters + agent executor
    pty/              -> PTY session module (spawn, Claude lifecycle, ansi)
  data/
    .agents/.library/ -> 20 pre-built agent templates
    getting-started/  -> Default KB page

Tech stack: Next.js 16, TypeScript, Tailwind CSS, shadcn/ui, Tiptap, Zustand, xterm.js, node-cron

Requirements

• Node.js 22+ (LTS). The repo ships an .nvmrc — run nvm use to auto-switch. Node 20 still works but produces an EBADENGINE warning from a transitive chevrotain@12 pulled in by mermaid.
• At least one supported CLI provider:
  • Claude Code CLI (npm install -g @anthropic-ai/claude-code)
  • Codex CLI (npm install -g @openai/codex or brew install --cask codex)
• Source mode: macOS, Linux, or Windows
• Electron desktop packaging: macOS and Windows

Configuration

cp .env.example .env.local

Authentication

Setting KB_PASSWORD turns on a single password gate for the whole UI/API (leave it empty for no auth). The session cookie is PBKDF2-HMAC-SHA256 over a per-install salt that's auto-generated into .cabinet.env on first run, the login endpoint is rate-limited against brute force, and the gate verifies in constant time. Changing the password (or salt/iterations) logs everyone out once. Full details, threat model, and tuning: docs/AUTH.md.

Commands

npm run dev          # Next.js dev server (port 4000 by default)
npm run dev:daemon   # Unified daemon: structured runs, terminal sessions, WebSockets, scheduler (port 4100 by default)
npm run dev:all      # Both servers
npm run electron:start   # Launch Electron desktop against the local dev servers
npm run build        # Production build
npm run start        # Production mode (both servers)
npm run elect