Lit Panel

Lit Panel is a local research workbench for reading, structuring, and connecting academic papers. It turns PDFs into reusable paper cards and research atoms, builds an interactive literature graph, and supports cited search, comparison, idea development, and AI-assisted synthesis.

The interface is bilingual (English and Chinese), while paper titles and source metadata remain in their original language.

Download

The first public desktop release is available from GitHub Releases.

• macOS Apple Silicon: download Lit-Panel-0.1.0-arm64.dmg
• The packaged app includes its own Node.js-compatible Electron runtime, portable Python runtime, backend dependencies, and a synthetic demo library.
• The current build is ad-hoc signed but not Apple-notarized. macOS may require opening it once through Finder's Open context menu.

The desktop release does not require a system Python or Node.js installation.

What It Does

• Imports local PDFs, NBER working-paper IDs, and DOI metadata.
• Runs configurable AI reading across research question, methods, data, identification, findings, mechanisms, limitations, and other dimensions.
• Lets the user independently enable or disable automatic Graph and Ideas updates for each reading run.
• Extracts reusable atoms such as methods, datasets, mechanisms, and puzzles.
• Provides FTS5 keyword search plus optional sentence-transformer semantic search, scoped to the active library.
• Builds formal paper, atom, and topic graphs without mixing unrelated search results into the selected topic.
• Supports paper scoring, notes, collections, comparison, cited Q&A, research maps, projects, and idea workspaces.
• Stores multiple local libraries with strict API-level library isolation.

Privacy Model

Lit Panel is a single-user local application. Its SQLite database, PDFs, notes, generated cards, graphs, and ideas stay in local storage unless the user exports them.

AI features are not fully offline: when the user starts an AI read, asks a question, or requests synthesis, the selected paper text and prompt are sent to the AI provider configured by the user. API keys are stored in the macOS Keychain when configured through the app and are used only to authenticate with that provider.

The public repository and packaged demo release exclude private databases, papers, .env files, API keys, project workspace paths, and build-machine bytecode caches.

Run From Source

Requirements

• macOS
• Python 3.10+ (3.11 recommended)
• Node.js 18.18+ (20+ recommended)
• An API key for Kimi, OpenAI, Anthropic, DeepSeek, Gemini, or MiniMax when using AI features

Start

git clone https://github.com/xtao-sh/litpanel.git
cd litpanel
./scripts/dev.sh start

Open http://127.0.0.1:3050. The backend runs at http://127.0.0.1:8050.

Command	Action
./scripts/dev.sh start	Install missing dependencies and start both services
./scripts/dev.sh stop	Stop both services
./scripts/dev.sh restart	Restart both services
./scripts/dev.sh status	Show service status

The first source run can take several minutes while Python dependencies and the embedding model are prepared.

First Run

1. Open Setup, choose an AI provider, add an API key, and enable it.
2. Open Pipeline and upload a PDF or enter an NBER ID / DOI.
3. Choose the reading dimensions and decide whether this run should update Graph, Ideas, both, or neither.
4. Select Start reading. The paper card and selected derived artifacts are written to the active library.

The downloadable desktop build starts with eight synthetic demo papers, twelve atoms, three ideas, and a small connected graph. To replace a source checkout's working database with the same synthetic seed:

backend/.venv/bin/python scripts/create_demo_db.py --force --replace-files

Back up a working database before using --force.

Configuration

The recommended configuration path is the in-app Setup page. For local development, an optional environment file is also supported:

cp backend/.env.example backend/.env

Data paths default to ./Data and ./backend/kb.db. Private data and local configuration are ignored by Git. See backend/.env.example for provider and path variables.

Architecture

• Backend: FastAPI, Strawberry GraphQL, SQLite/FTS5, optional semantic embeddings, and the multi-step reader/linker pipeline under agents/.
• Frontend: Next.js App Router with the Lit Panel design framework, library, paper, pipeline, graph, atlas, ideas, maps, projects, and setup views.
• Desktop: Electron starts the packaged FastAPI and standalone Next.js services on loopback ports 38000 and 38001, then opens the local UI.

See INTRODUCTION.md for additional architecture notes and desktop-mvp/README.md for desktop build details.

Validation

cd frontend && npm run lint && npm run build
cd ../backend && ./.venv/bin/python -m unittest discover -s tests -v
cd ../desktop-mvp && npm audit && npm run smoke

Release builds are additionally checked for clean demo data, absent API keys and project workspace paths, process shutdown, code-signature integrity, and DMG checksums.

Limits

• The downloadable release currently targets macOS Apple Silicon (arm64).
• Public multi-user deployment is not supported; the API is designed for a local single-user workspace.
• Semantic search needs the embedding model. Keyword search remains available without it.
• Public distribution without Gatekeeper warnings requires an Apple Developer certificate and notarization.

License

MIT - xtao-sh