📰 The GPT Newspaper

Your own newspaper, researched, written, fact-checked and illustrated by a newsroom of AI agents.

Tell it what you care about. Reporters search the news with Tavily, editors argue over every draft, an art director commissions illustrations, and the editor-in-chief hands you a front page in about a minute. Every claim is cited.

✨ Highlights

• Any model. One provider:model string through LangChain: OpenAI, Anthropic, Google, Groq, Mistral, xAI, DeepSeek, or a local model with Ollama. Pick a default on the server or switch per edition.
• A real newsroom, not one prompt. An assignment editor plans the coverage, reporters search in parallel, a curator throws out off-topic results, writers draft from full article text, a critic sends drafts back for revision, and an art director commissions the art.
• Cited, grounded journalism. Inline citations on every factual claim link to the numbered sources. Sources are treated as untrusted data, never as instructions.
• Illustrations in a house style. Ink, Engraving or Risograph art in the paper's own palette, sized for each layout slot and always labeled "AI illustration". Or use the sources' news photos.
• Watch it work. The newsroom streams live: every topic, every stage, the editor's notes to the writer, and stories landing on a live front page as they finish.
• Resilient. If a topic has no news this week, reporters dig further back. If a story fails, the rest of the edition still prints. Transient API errors are retried.
• 17 languages, right-to-left included. Print or save as PDF with a proper print stylesheet.
• Bring your own keys in the UI (they stay in your browser), or configure them on the server.
• Web app, API, CLI and LangGraph Studio, from one codebase.

🚀 Quick start

You need Python 3.11+, a free Tavily API key, and a key for your model provider (or a local Ollama).

git clone https://github.com/rotemweiss57/gpt-newspaper.git
cd gpt-newspaper
pip install -r requirements.txt
python app.py

Open http://localhost:8000, add your keys under Model & API keys, pick some topics, and press Print my newspaper.

With uv

uv sync --extra anthropic   # extras: anthropic, google, ollama, groq, mistral, all
uv run app.py

With Docker

docker compose up --build

Keys can go in a .env file (see .env.example) or in the web UI. Editions are saved to ./outputs.

🗞️ How the newsroom works

Everything is a LangGraph StateGraph: topics fan out with the Send API, per-run settings travel in the runtime context, and progress is streamed with the stream writer. Every LLM call is a LangChain runnable with a typed structured output, so swapping models is a one-line change.

🧠 Models

Set LLM_MODEL (or use the Model field in the UI) to anything LangChain's init_chat_model supports, and install that provider's integration:

Illustrations are drawn by IMAGE_MODEL (default openai:gpt-image-2) and need an OpenAI key, even when the articles are written by another provider. Without one, editions use the sources' news photos.

⚙️ Configuration

All settings are optional environment variables (or a .env file). See .env.example.

⌨️ CLI

Print an edition from your terminal, or from cron for a paper every morning:

python -m backend.cli "AI agents" "Formula 1" --layout front --images engraving --language es --open

🔌 API

POST /api/newspaper streams newline-delimited JSON while the newsroom works:

curl -N localhost:8000/api/newspaper -H 'Content-Type: application/json' -d '{
  "topics": ["AI agents", "Space exploration"],
  "layout": "layout_3",
  "images": "ink",
  "language": "en",
  "model": "anthropic:claude-sonnet-5-5"
}'

{"type": "progress", "topic": "AI agents", "index": 0, "stage": "plan", "message": "Assigned: “AI agent funding” · “AI agent security”"}
{"type": "progress", "topic": "AI agents", "index": 0, "stage": "review", "message": "Editor: Lead with the funding news."}
{"type": "progress", "topic": "AI agents", "index": 0, "stage": "done", "message": "On the front page", "article": {"title": "…", "url": "…"}}
{"type": "done", "url": "/outputs/run_1760000000_ab12cd/newspaper.html", "articles": 2, "failed": 0}

Optional fields: api_key (model provider), tavily_api_key, and image_api_key override the server's keys for that request. Layouts: layout_1 (Classic), layout_2 (Digest), layout_3 (Front page). Also available: GET /api/editions (archive), GET /api/config, GET /api/health, and the v1 POST /generate_newspaper.

🧪 LangGraph Studio

pip install "langgraph-cli[inmem]"
langgraph dev

Run the newspaper graph with {"topics": ["AI agents"], "layout": "layout_1"} and watch every agent.

🗂️ Project layout

app.py                 # starts the web server
backend/
  graph.py             # the LangGraph newsroom and the streaming runner
  agents/              # one module per agent
  prompts.py           # LangChain prompt templates
  schemas.py           # Pydantic structured outputs
  images.py            # illustration styles, slot sizes, image provider
  setup.py             # request → validated run context (shared by API and CLI)
  server.py            # FastAPI: UI, streaming API, archive, security headers
  cli.py               # command-line interface
  templates/           # Jinja2 newspaper layouts and article page
frontend/              # the web app: plain HTML, CSS and JS, no build step
tests/                 # the full graph and API, against a fake model, search and image provider

🤝 Contributing

Contributions are very welcome. Read the contributing guide, and run the checks before opening a PR:

uv sync && uv run pytest && uv run ruff check . && uv run ruff format --check .

Found a security issue? Please follow the security policy.

🛡️ Disclaimer

GPT Newspaper is an experimental project, provided "as is". Articles are generated by AI from web sources and can contain mistakes: check the cited sources before relying on anything. Illustrations are AI-generated and labeled as such. It's meant for personal use, not as a replacement for professional journalism.

📩 Contact

rotem5707@gmail.com · MIT licensed