Kanbini

A kanban board that lives on your machine - and lets your AI move the cards. Offline-first. Plain text on disk. No cloud, no accounts, no telemetry.

Kanbini is a Trello-style kanban app built for one person and their AI. Your boards live in a local SQLite file and export to plain Markdown + JSON, so you own your data outright. A built-in local MCP server lets any AI that speaks MCP - Claude, ChatGPT, Cursor, or your own agent - read and rearrange your board while you watch the cards move in real time.

The pitch is simple: your tasks shouldn't need an internet connection, a login, or a monthly fee just to exist. Kanbini keeps your whole board in a local file you fully own, runs as a fast native app, and - when you want a hand - lets an AI pull up a chair and move the cards in real time while you watch.

The name mashes kanban with konbini, the Japanese corner store that never closes: a board that's yours, local, and always on.

Why Kanbini

• Offline by default. No accounts, no network, no telemetry. The only things that ever leave your machine are the ones you explicitly turn on
  • link previews and one-way Obsidian export - both off by default.
• You own your data. Everything is a local SQLite file with first-class, lossless plain-text export (Markdown + JSON), verified byte-identical by a round-trip test.
• AI-native. A token-gated local MCP server exposes read/write tools; the AI's edits stream into the live UI as it works.
• Free and open source under the MIT license.

Screenshots

All your boards in one place	Card detail with checklists + comments
Swimlanes - group by priority	Cross-board command palette

Features

• Boards, lists, cards with drag-and-drop (live cross-list reorder), labels, due dates, priorities, per-list colours, and swimlanes (group by priority).
• Card depth - WYSIWYG description (saved as Markdown), checklists, comments, local attachments (paste an image straight in), cover images, and a per-card activity log. Cards with a description show a small marker in the list.
• Multi-select - Ctrl/Cmd-click cards (or Shift-click for a range) to pick several at once, then complete, set priority, re-label, move, or delete them in bulk from a floating action bar (or a right-click menu) - and drag the whole group between lists.
• Local MCP server - a token-gated 127.0.0.1 control channel + stdio server (@kanbini/mcp) with read/write tools (get / create / update / move / delete / label / comment / checklist / search). Edits stream into the open UI live. See docs/MCP.md.
• Scriptable HTTP API - the same token-gated 127.0.0.1 control channel is also a plain local HTTP API (REST routes for reads + writes, plus a batch endpoint that lands as one undo step) for non-MCP tools, scripts, and automation. Loopback only. See docs/MCP.md.
• Own your data - lossless plain-text export/import (kanbini.json + cards/*.md + attachment files); Backup + Restore live in Settings.
• Cross-board search via a Ctrl/Cmd+F command palette.
• Templates for boards and lists.
• Undo / redo with full snapshot restore on delete.
• Customizable keyboard shortcuts.
• Board backgrounds (colour / gradient / image) and continuous zoom.
• Trello import - bring an exported Trello board JSON in as a new board.
• Opt-in integrations (off by default): link previews and one-way push to an Obsidian vault.

Install

There's no published installer yet - run it from source:

pnpm install
pnpm --filter @kanbini/desktop dev

It opens on a seeded sample board. After a fresh clone or any Electron version bump, run the native rebuild once:

pnpm --filter @kanbini/desktop rebuild:native

To build a real installer, see docs/PACKAGING.md (Windows NSIS + portable .exe; macOS .dmg; and a Linux AppImage, built from a Linux host or WSL2 - see Building for Linux). Windows is the primary tested target; the macOS and Linux recipes build from the same Electron source but haven't been smoke-tested yet.

Wire up an AI

The MCP server speaks the standard Model Context Protocol, so any MCP-capable client can drive your board while it's open. Claude Desktop and Claude Code are what we test against, but ChatGPT desktop, Cursor, or your own agent work the same way - one config shape covers them all. The snippet + token live in the app under Settings → AI integration; details in docs/MCP.md.

Develop

pnpm verify   # typecheck + unit tests + launch smoke (~20 s)
pnpm e2e      # Playwright end-to-end suite against the real Electron app

Exercise the MCP surface end-to-end (desktop app must be running):

pnpm --filter @kanbini/mcp run build         # one-time / after MCP changes
pnpm --filter @kanbini/mcp run smoke         # read-tool check
pnpm --filter @kanbini/mcp run smoke:write   # create → update → comment → checklist → delete

Prove the export/import round-trip is lossless (no live app needed):

pnpm --filter @kanbini/desktop run test:roundtrip

Layout

/apps
  /desktop      Electron main + hardened preload + control channel
  /renderer     React + Vite + Tailwind v4 + ShadCN
  /mcp          @kanbini/mcp - stdio MCP server
  /desktop-e2e  Playwright end-to-end suite
/packages
  /shared       zod schemas, channel names, id/order utils
  /db           Drizzle schema + migrations + data access
/docs           DESIGN.md, MCP.md, PACKAGING.md

• Design & architecture → docs/DESIGN.md
• MCP integration → docs/MCP.md
• Packaging → docs/PACKAGING.md

Contributing

Issues and pull requests are welcome. pnpm verify should be green before you push. New features in @kanbini/shared, @kanbini/db, or @kanbini/renderer should land with tests in the same change; bug fixes should come with a regression test.

License

MIT © Jigglyy and Kanbini contributors.