● open source · mit · terminal task manager for the agent era
Your agents
got an office.
taskdeck discovers the work scattered across your project board, Claude Code sessions, markdown checklists and git commits — then dispatches an agent to any task and shows it working. In your terminal, and in a pixel office where every project is a room.
npx taskdeck
Zero config · Node 18+ · runs in any git repo
Most to-do tools make you enter work. taskdeck discovers it — folding four sources you already produce into one filterable, keyboard-driven list:
.taskdeck/board.json) that persists across every chat and session.- [ ] markdown todos in your repo's docs.GitHub, Linear and Notion issues plug in as extra sources.
Open taskdeck in any repo. Board, sessions, checklists and commits are already there — filter by source or status, fuzzy-search everything.
Select a task, press a. Choose investigate (read-only plan mode) or execute (the agent does the work). A parallel-run queue caps how many agents work at once.
Progress streams live in the Runs view — and the agent appears in its project's room, typing at a desk. Every task shows the tokens its runs consumed, so cost is never a surprise.
Finished runs notify your desktop with a concise report. Press c to resume the agent's session and keep going.
taskdeck Tasks Projects (3) Runs (1 active) source all 409 board 57 sessions 323 docs 4 commits 25 status open 46 done 363 ──────────────────────────────────────────────────────────────────────────── ▸ ◐ board Refactor the config layer to be repo-agnostic Ship taskdeck ○ board Publish v0.1 to npm Ship taskdeck ● session Add GitHub issues as a task source Build sources ● commit feat: parallel run queue + notifications a1825c3 ○ doc Write the keybindings table README.md ──────────────────────────────────────────────────────────────────────────── ↑↓ nav · ⏎ details · a run agent · / search · s source · f status · r runs · q quit
Built with Ink — React for CLIs. taskdeck all shows every project's board in one view.
When an agent hits a permission it can't grant itself, it doesn't fail silently — it waits in the lounge with a speech bubble. Approve in its terminal and it walks back to its desk.
And you don't have to be at your desk either: set an ntfy.sh topic and finished runs push to your phone with the run's report and a Continue button that resumes the agent's session — exactly like pressing c in the TUI. No server, no LAN setup.
MIT licensed, local-first, zero servers. Your board is a JSON file in your
repo; your transcripts never leave your machine. Agents are plain
claude -p subprocesses — bring your own
Claude Code.