Mission control for Claude Code. One localhost app that watches every project on your machine, every agent you're running, and everything they're doing — live.
If you run multiple Claude Code sessions in parallel, you know the ritual: eight terminal tabs, alt-tabbing to see who's stuck, a GUI git client to review what the agents actually changed, and a sticky note for what you were supposed to be doing. Deck replaces the ritual with one screen.
Everything here runs on
127.0.0.1. Deck never opens a port to the network, never phones home, and reads your Claude Code transcripts straight off disk.
Every agent, one sidebar. Deck discovers every git repo under your code root and every Claude Code session on your machine — including ones you started in a plain terminal, which it tails read-only from their transcripts. Each project card shows its live agents with status (working / waiting on you / finished / exited), the original prompt you gave each one, and what it's doing right now. When an agent stops to ask a question, the exact prompt it's stuck on shows up in the sidebar with a quick-reply box.
A transcript feed that reads like the terminal. The agent feed renders the way Claude Code itself does — monospace, ● tool bullets, ⎿ result branches, expandable mini-diffs for every edit. Live sessions stream in over WebSockets; a 1,000+ event transcript scrolls fully virtualized.
Real terminals, not toys. Spawn claude or plain shells in server-side PTYs (ConPTY), rendered in xterm.js with the WebGL renderer. Native-speed typing, correct resize with no garbling, and reattach that restores your exact screen after a browser refresh — or a server restart.
Git that keeps up with agents. Status, word-level diffs, byte-exact hunk staging, commit/push, log and commit viewer — plus an AI PR Audit: a pre-merge report with a risk level, bug/risk/nit findings that jump to the diff, impact areas, and a before-merge checklist. There's an "ask about this change" box when you want a second opinion on the second opinion.
A task board that minds its own business. A personal kanban — Inbox / Next / Now / Done — built for dump-first capture: paste a screenshot, hit Enter, assign a project later. It has no due dates, and it will never start an agent behind your back. Its one piece of automation: ✨ drafts a ready-to-paste Claude prompt from a card plus the project's docs.
A library, not a list. 150 repos is a junk drawer. Deck groups them into bands you curate, enriches each card with a blurb, stack badges and run buttons, detects live dev servers by port, and shelves whatever you haven't touched in a month.
And the long tail: full-text search across every transcript on your machine (SQLite FTS5) · a review queue of files agents changed while you weren't looking · AI tab titles and one-line summaries for every live session · a daily digest of what shipped · cost dashboards with per-feature budgets · an embedded app preview tab · a system view for ports and orphaned processes · env-file and Postgres inspection with a built-in read-only query box.
Deck's own AI features (titles, summaries, audits, digests) route through a single choke point with per-feature daily budgets and an append-only usage ledger, and default to cheap models. Two backends:
claudeCLI (default) — uses the Claude Code binary you already have. No API key.- Anthropic API — set
anthropicApiKeyin config if you prefer.
Blow a budget and features degrade gracefully (cards keep their last title) instead of silently spending. Turn the whole layer off in AI Admin and Deck still does everything non-generative. Your .env values are never sent to any model.
The server survives uncaught exceptions (logged to ~/.deck/crash.log), and if something truly kills it, a supervisor restarts it with backoff in under 5 seconds. Terminal scrollback flushes to disk every 30s, state writes keep a .bak, and restored tabs reconnect to what's on disk — a hard kill loses nothing.
- Windows 10/11 — Deck drives real ConPTY terminals and leans on PowerShell for system introspection. macOS/Linux isn't supported yet (the pty layer and the system view are the two Windows-specific spots; PRs welcome).
- Bun ≥ 1.1 — package manager + script runner
- Node.js ≥ 22 — the server runs on Node (via tsx); node-pty needs it
- Claude Code CLI on your
PATH(where claudeshould hit) - git on your
PATH - PowerShell 7 (
pwsh) for spawned shells — or pointdefaultShellatpowershell.exe
git clone https://github.com/dested/deck.git
cd deck
bun install
# development (hot reload) — server on :12345, web on :12346
bun run dev # open http://127.0.0.1:12346
# production — one origin, supervised auto-restart
bun run build
bun start # open http://127.0.0.1:12345First run: open Settings (or create deck.config.json, below) and point Deck at your code folder. Every direct child with a .git is picked up automatically; Claude sessions appear as soon as they write a transcript.
deck.cmd (double-click, or pin to the taskbar) starts the prod server if needed and opens Deck in a chromeless Edge/Chrome app window with its own taskbar icon. Deck is also installable as a PWA from the browser menu.
Optional deck.config.json at the repo root:
Deck keeps its own state in ~/.deck/ (UI state, search index, usage ledger, crash log, scrollback). It reads ~/.claude/projects/ read-only — it never writes to or moves your transcripts.
bun run typecheck # tsc --noEmit across server + web
bun run --filter @deck/server start:unsupervised # server without the supervisorThree workspaces: shared (every server↔client type, one file), server (Fastify + WebSockets + node-pty + chokidar, Node 22), web (React 19 + Vite + Tailwind v4 + xterm.js + Monaco).
The heartbeat is the filesystem: chokidar tiers watch your code roots, ~/.claude/projects, and each repo's .git/index — so a file an agent touches bumps its project within ~2 seconds, without Deck ever recursively statting 150 worktrees. Transcripts are parsed tolerantly (unknown line types can never crash the feed), diffed live, and streamed to subscribed clients over /ws/events; terminals ride a binary bridge on /ws/term with server-side screen serialization for perfect reattach.
The frontend self-hosts everything — fonts, Monaco, icons. Deck works completely offline (minus the AI features, obviously).
Deck itself was built agent-first: a locked spec (SPEC.md, SPEC2.md) handed to Claude Code milestone-by-milestone, with the kickoff prompts preserved in KICKOFF.md and a living map in cliffnotes.md. The repo is its own proof of concept — Deck was built while being dogfooded to watch itself being built.
Why localhost web and not Electron? A server has to exist anyway (PTYs, watchers, transcript parsing), and a browser tab you can also install as a PWA is lighter than shipping Chromium twice. The deck.cmd app window gets you the desktop feel for free.
Does it work with agents I start outside Deck? Yes — that's half the point. Anything writing a transcript to ~/.claude/projects shows up live, read-only, with an Adopt button that resumes it inside a Deck terminal (claude --resume).
Port 12345 is taken / a dead server is squatting it.
Get-NetTCPConnection -LocalPort 12345 -State Listen | % { Stop-Process -Id $_.OwningProcess -Force }Terminal icons render as □? They shouldn't — Deck bundles a symbols-only Nerd Font fallback for oh-my-posh / eza glyphs. If you see tofu, hard-refresh so the font loads.
How much do the AI features cost me? Default budgets cap Deck's own spend at a few dollars/day across all features, most of it on the cheapest model tier, and the AI Admin view shows a full per-call ledger. Zero if you turn the layer off.
Issues and PRs welcome. The fastest way to orient is cliffnotes.md (the living map of the codebase), then SPEC.md for the original design intent. bun run typecheck must stay green in both packages.
{ "root": "D:\\code", // primary projects folder (default: ~/code) "roots": ["D:\\work"], // extra folders to scan (optional) "port": 12345, // server port "defaultShell": "pwsh.exe", // shell for new terminals "claudeBin": null, // override `claude` resolution "claudeDir": null, // override ~/.claude "anthropicApiKey": null, // use the API instead of the claude CLI "digestAt": "18:00", // daily digest time (omit to disable) "webstormBin": null // for the "open in WebStorm" button }