Kiwi is not another note-taking app that stops at storing things.
Documents, PDFs, web pages, spreadsheets, canvases, and notebooks are raw material here. A compilation pipeline breaks them into concepts, merges them into cross-referenced Wiki pages, and preserves the source behind every conclusion. The same knowledge then enters the learning engine—profile, practice, spaced repetition, mock exams—and mastery is written back to the pages.
Most tools stop at “captured.” Kiwi keeps asking: How do you know you have actually learned it?
Why it is different
- Knowledge is compiled. No hand-maintained hierarchy. Schema, incremental compilation, and references let structure grow on its own.
- The graph is also a mastery map. Page references, note links, weak spots, and mastery share one coordinate system—see how knowledge connects and where you have not learned it yet.
- The learning loop lives inside the knowledge base. Questions, review, exams, and mastery updates all happen in the same library; no exporting content to another tool.
- Local-first, swappable models. Runs on a single machine with PostgreSQL + Qdrant and supports any OpenAI-compatible model; data and inference stay on your hardware.
Core capabilities
- Note workspace: docs, databases, sheets, imports, links, canvas, and notebooks share folders, tags, templates, and bidirectional links.
- Two editing modes: switch the same Markdown between the standard editor and DOCX—block-based writing, interactive blocks, and formal layout in one document.
- Theme plugins: import declarative theme packages with color tokens, background images, panel opacity, glass blur, and typography; the built-in Classic and Aurora themes restyle the editor, canvas, tables, and component plugins together.
- Executable code blocks: Python runs in a bubblewrap sandbox, with output and exit code kept in the document; HTML / JS previews inside an isolated iframe.
- Advanced capabilities: voice transcription, sticky note windows, component plugins, read-only sharing, sheets with OCR, hybrid retrieval QA, todo hub, and typed exports.
- Compilation jobs: every eligible note compiles into the Wiki—todo hubs, read-only imports, and database row notes are skipped—with job progress and affected pages tracked.
- Structured Wiki: schema-driven, incrementally updated, source-cited, with health checks and one-click rebuild.
- Knowledge star map: explore page relationships as a star map; node size reflects citation count.
- Learning loop: profile → plan → progressive practice → spaced repetition → mock exam → mistake notebook.
- Answers with sources: retrieval across the Wiki and notes, with persistent chat sessions and inline citations; answers are traceable and can be saved as new pages.
Interface
Tech stack
Vue 3.5 · TypeScript · Vite · Pinia · Tailwind CSS · TipTap · AntV G6
FastAPI · SQLAlchemy 2 · PostgreSQL 16 · Qdrant · fastembed
Docker Compose · uv · Node.js
Getting started
Requires Python 3.12, uv, Node.js ≥ 20, and Docker:
git clone https://github.com/LsWGG/Kiwi.git
cd Kiwi
./run.sh
http://localhost:5173 opens the product portal by default; the app workspace is at http://localhost:5173/kbs. Add any OpenAI-compatible model in Settings to start. If local embedding models cannot load in your environment, switch to a hosted embedding API under Settings → Embedding model.
On first launch, Kiwi seeds the built-in “Kiwi Quick Start · Feature Demo” knowledge base, containing 41 notes, 34 Wiki pages, and complete learning and exam data. You can edit, delete, or restore it at any time.
Windows desktop (optional)
A standalone Windows x64 build wraps the same app in a WebView2 window and bundles PostgreSQL, Qdrant, and a portable Python runtime. No LLM, embedding, Whisper, or OCR models are included; all data stays local by default and AI features activate only after you configure a local or OpenAI-compatible endpoint. Docker remains the primary deployment.
Download Kiwi-Setup-<version>.exe from Releases, or build it on Windows 10/11 x64:
powershell -ExecutionPolicy Bypass -File desktop\build.ps1 -Version 1.1.1
See docs/WINDOWS.md for installation, offline boundaries, local code execution, backup/restore, and troubleshooting.
Docker deployment
Production uses a separate Compose file with non-standard ports by default: Web 18473, API 18474, PostgreSQL 15432, Qdrant 16333 / 16334.
The frontend is built on the host and mounted read-only into nginx — the Docker image no longer runs npm or Vite:
(cd frontend && npm ci && npm run build)
./deploy.sh
./deploy.sh starts a step-by-step wizard (browser origin, access token, frontend rebuild, backup, and optional Cloudflare Tunnel), then builds the backend image and runs database migrations. Use ./deploy.sh --yes for non-interactive deployments, ./deploy.sh frontend to rebuild the static bundle only, and ./deploy.sh cloudflare to configure public read-only sharing. Rebuild frontend/dist and refresh the browser after frontend changes — no container rebuild is needed. Then open http://<host>:18473.
See docs/DEPLOY.md for port changes, backups, upgrades, sandbox permissions, and troubleshooting. See the changelog for releases and the roadmap for what comes next.
Tests
# Backend: pure logic + API contract tests (real PostgreSQL / Qdrant)
cd backend && uv run pytest -q
# Frontend type check and Playwright smoke tests
cd frontend && npm run type-check && npm run test:e2e
Playwright expects a backend at http://127.0.0.1:8000 using an isolated test database with the background worker disabled (KIWI_BACKGROUND_JOBS=false, KIWI_RATE_LIMIT_ENABLED=false). Set PW_TOKEN_BACKEND_URL and PW_ACCESS_TOKEN to also run the access-control login flow. CI runs the full suite on every push and pull request.
License
Copyright © 2026 LsWGG.
Kiwi is free software: you can redistribute it and/or modify it under the terms of the GNU Affero General Public License, version 3 (SPDX: AGPL-3.0-only), as published by the Free Software Foundation.
Kiwi is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See LICENSE for the full license text.
The distributed application bundles PyMuPDF under AGPL-3.0, so the combined work is provided under the AGPL-3.0. The corresponding source is available at https://github.com/LsWGG/Kiwi; third-party notices are listed in THIRD-PARTY-NOTICES.md.
