Put your coding agent on a kanban board.
Drop your tasks on the board, hit run, and Claude Code, OpenCode or Pi picks one up -- in its project directory, already briefed on the task. You follow along as a readable conversation or in the real terminal embedded in the page, and the agent's report waits beside it for your sign-off. A queue works through the rest unattended, one task per project, while you do something else. Runs on your own machine: one Python package, a SQLite file, no account, no build step, nothing leaves the box.
uv tool install ntasker # install from PyPI
ntasker service install --auto-update # run as a service + daily auto-update
ntasker config set projects_dir ~/ProjekteOpen http://127.0.0.1:8766. The service creates the database on first start, restarts on crash and keeps itself up to
date. On Linux, run loginctl enable-linger $USER once so it survives logout. No supervisor wanted? ntasker serve
runs it in the foreground until you close it.
Then teach your agent about it:
ntasker agent install claude # or: opencode, piThree things a plain to-do list cannot do.
The installed skill and /task command let the agent read and drive the tracker -- no copy-paste, no re-explaining:
- "What should I work on next?" -- it grabs the open tasks for your current project folder and ranks them by urgency.
/task 34-- pulls #34 into the session (title, description, tags), flips it to in progress, and warns you if you are sitting in the wrong project.- "Add a todo: ..." -- it files the task for you; drop a
#34anywhere later and it knows which task you mean. - Finished an assigned task? It moves the task to Review for you to sign off. It never closes, deletes or archives anything on its own.
Every task row has a run button showing that task's agent logo. It starts the genuine agent CLI in the task's
project directory, seeded with the task -- same CLI, same CLAUDE.md, same skills, MCP and permissions. The run view
offers four ways to look at it:
- Conversation -- the session transcript folded into readable turns: your prompt, the agent's answer, the steps and tool calls behind it, the tokens it used. Permission prompts and questions get buttons, and a reply box sends the next prompt. If the agent cannot work (usage limit, not logged in, billing, API unreachable), a blocker card says why and what to do.
- Terminal -- the real TUI via xterm.js: answer, approve and interrupt exactly as in a terminal.
- Split -- both side by side. Diff -- what the run changed in the working tree.
- Report -- the agent's final report in its own pane with a Mark done button. It opens by itself when the report is written and warns when the conversation went on after it.
Sessions run in the background (the button shows a spinner; re-opening reattaches to the live session). Pause ends a session resumably, marking a task done ends it for good. An open task whose session already ended gets a resume button next to Run, which continues that conversation instead of starting over. With a claude.ai subscription, the topbar shows the 5-hour and weekly usage limits as meters. The conversation view needs a readable transcript (Claude Code); OpenCode and Pi get the terminal. See docs/claude-runs.md.
Every run button puts its task at the head of its project's lane, and ntasker works through the queue one task per
project at a time, taking the next one as soon as the previous is closed -- by you after review, or by the agent when
the task told it to. done is the only thing that ends a session. Pause stops new starts; running tasks keep going.
The panel shows one column per project, because that is what runs in parallel. To make one task wait for another -- across projects too -- drop it on the middle of the other; the edges keep reordering. A fasttrack task commits and closes itself and hands its result to the tasks depending on it; the Run-Log collects those outcomes. Plan queue lets an agent session order the queue for you and start it. See docs/task-queue.md.
Type a raw thought into the topbar field (Ctrl+K) or ntasker in "..." and carry on. A stateless claude -p call
turns it into a proposed task -- title, prompt, priority, tags and the project, chosen from a catalog of one-paragraph
project summaries -- and the inbox list above the queue shows the proposal with its candidate projects pre-ticked. One
click accepts it (or accepts and runs it); tick another project, or several (the rest become directory locks), or
answer the proposal to re-triage it. Every correction becomes an example for the next triage. Nothing becomes a task
until you say so.
See docs/inbox.md.
The sidebar lists every project under your projects_dir as a tree of families and groups, with a search field by
name or group. The New project dialog creates one on disk -- directory under the projects base, optional
git init, sidebar group -- and can start an agent in it right away. Conversations you began in a plain terminal can
be filed under a task later (ntasker adopt or Pick up a session) and resumed from the board. See
docs/projects.md.
ntasker is agent-agnostic: Claude Code, OpenCode and Pi are supported out of the box, and adding another is one
plugin. Each task carries an agent and an optional model; either can fall back to a global default.
ntasker add --title "..." --agent opencode --model opus
ntasker config set default_agent opencode
ntasker agent list # CLI availability + integration status per agentFull reference incl. per-agent binary paths: docs/agents.md.
Optional features ship as plugins you can switch off individually (/settings -> Plugins):
| Plugin | What it adds |
|---|---|
task_context |
Attach files, notes, personas, skills and MCP servers to a task, handed to the agent in its briefing (docs) |
workspace |
Team (Claude Code subagents), skills, knowledge base and documents on a /workspace page (docs) |
voice |
Dictate task descriptions with local speech recognition; opt-in via ntasker enable voice (docs) |
A disabled plugin's routes 404, its agent is neither listed nor resolvable, and its data stays intact. Contract and slots: docs/plugins.md.
| Topic | |
|---|---|
| Configuration | Settings, DB path, projects_dir, language, vendor assets, why you must not expose the port |
| CLI reference | Every subcommand and flag |
| HTTP API | Endpoints, SQLite schema, design notes |
| Agents | The agent registry, per-task model, skill installation |
| Agent runs | Run view (conversation, terminal, diff, report), resume, usage meters |
| Task queue | Queue semantics, dependencies, fasttrack, run log |
| Directory locks | Keeping two agents out of the same working directory |
| Kanban view | Board vs. list view, drag-and-drop, keyboard shortcuts |
| Inbox | Raw notes triaged into task proposals by a stateless claude -p call |
| Projects | Sidebar tree, project families, search, new-project dialog, misc project |
| Service | systemd / launchd, auto-update, uninstall |
| Development | Repo setup, smoke test, translations |
Coming from the drfoehn fork? See docs/migrating-from-fork.md.
FastAPI + uvicorn on the Python stdlib sqlite3 -- no ORM, no migration files. The frontend is HTML + AlpineJS +
Tabler.io loaded from jsDelivr with pinned SRI hashes, so the wheel stays under 100 KB and there is no build step; an
offline mode is one command away. Requires Python 3.12+.
Binds to 127.0.0.1:8766 and has no authentication -- it is a personal local tool, not a multi-user service.
AGPL-3.0-or-later. The Affero clause means: if you run a modified version as a network service, you must offer the modified source to its users. For local single-user use this has no practical impact.
Changelog: CHANGELOG.md -- issues and source: https://github.com/nerdocs/ntasker
If it saves you an afternoon, you can buy me a coffee.