Your own daily news brief, written by Claude on your Claude subscription, from the sources you choose. NewsFramer reads your RSS feeds every morning, drops duplicates and noise, scores each story against the topics you care about, and writes one themed brief with every source cited. You get it as a Markdown and JSON file to publish on your site, paste into a newsletter, or send to your phone.

NewsFramerは、自分で選んだRSSを毎朝読み、Claude Code経由の自分のClaude Pro/Maxで日次ニュースブリーフを書く個人用ツールです。 重複とノイズを除き、関心のあるトピックごとに記事を採点して、出典つきのテーマ別ブリーフにまとめます。出力はMarkdownとJSONなので、自分のサイトに公開したり、ニュースレターに貼ったり、スマホへ送ったりできます。

A real brief · Quickstart · Guide for AI coding agents · (logic tests on Linux and Windows)

Easiest path: open this repo in Claude Code or Codex and say "Set up NewsFramer for me by following README.md". The agent section below tells it everything it needs.

By hand (PowerShell shown; on macOS or Linux use python3 -m venv venv and source venv/bin/activate):

py -3.13 -m venv venv; .\venv\Scripts\Activate.ps1
pip install -r requirements.txt
npm install -g @anthropic-ai/claude-code   # then run `claude` once and sign in with your Claude plan
python setup_wizard.py                     # creates your config files and .env (asks for 2 Supabase values)
python manage.py db-init                   # creates the database tables
python manage.py sync                      # loads your sources and interests
python setup_check.py                      # PASS/FAIL for every part, including a live Claude call
python run_daily.py                        # builds today's brief into output/brief-<date>.md

You need: Python 3.13, a Claude Pro or Max plan with Claude Code, and a Supabase project (its smallest plan is enough). Everything else is optional.

This section is written for Claude Code, Codex and similar agents setting up, adapting or extending NewsFramer for a user. Follow it literally.

A daily batch pipeline in Python 3.13. Each step is a standalone script under agents/; steps never call each other, Supabase (Postgres + pgvector) carries state between them, so any step can be re-run alone.

run_brief.py runs the five steps; run_daily.py runs them and then deliver_brief.py. All AI calls go through headless Claude Code (claude -p, see agents/cc_writer.py); an API model is used only as a fallback, and every fallback is printed, logged to logs/subscription-fallback.log and alerted.

1. Python 3.13 venv, pip install -r requirements.txt.
2. Claude Code installed (claude --version) and signed in by the user (claude, interactive; ask the user to do this step).
3. python setup_wizard.py creates config/models.yaml, config/sources.yaml, config/interests.yaml from the *.example.yaml templates and writes .env. Required in .env: SUPABASE_URL, SUPABASE_SERVICE_KEY. Optional: SUPABASE_DB_URL (for db-init; in Supabase: Connect > Connection string; use the Session pooler URI if the direct one fails, since the direct host is IPv6-only), GEMINI_API_KEY (dedup + fallback), ANTHROPIC_API_KEY (fallback), TELEGRAM_BOT_TOKEN + TELEGRAM_CHAT_ID (alerts), DELIVERY_TARGET, FIRECRAWL_API_KEY, OPENCLAW_MJS. Never ask the user to paste a key into chat; they type it into the wizard or .env.
4. Tables: python manage.py db-init (needs SUPABASE_DB_URL), or paste sql/schema.sql into the Supabase SQL editor. Safe to re-run.
5. Set operator_timezone in config/models.yaml (IANA name).
6. Edit config/sources.yaml and config/interests.yaml to the user's topics, then python manage.py sync.
7. python setup_check.py must print all PASS.
8. python run_daily.py, then open output/brief-<date>.md and confirm it has themes and cited links.
9. Schedule python run_daily.py daily (Task Scheduler, cron, or an OpenClaw cron).

• Config only. Every number lives in config/models.yaml, read with config.get(key, default). Never put a tunable in code or in a prompt file. A new key goes into config/models.example.yaml too.
• Subscription first. Do not switch steps to a metered API by default. Fallbacks must stay loud (cc_writer.notify_fallback).
• Soft-delete only. Articles are never hard-deleted; dedup sets deleted_at.
• Delivery is recorded only after a confirmed send (agents/deliver.py), or after the file is saved in file-only mode. This is how tomorrow's brief knows what is new.
• Facts come from articles. The writer summarises source articles with citations; it never rewrites a previous brief.
• Secrets stay in .env. Never print, log or commit a key. git config core.hooksPath hooks turns on the pre-commit key guard.
• Tests: python tests/run_all.py (no database or AI calls) must stay green; add a test with every change.

run_daily.py            build + save + deliver (the daily job)
run_brief.py            build only (the five steps)
deliver_brief.py        save .md/.json, post_brief_command, optional OpenClaw send
run_whatsapp_brief.py   per-chat WhatsApp briefs
manage.py               db-init | sync | sources
setup_wizard.py         first-run config + .env;   setup_check.py  the install check
check_run_health.py     missed-run watchdog;       gateway_watchdog.py  OpenClaw gateway watchdog (Windows)
agents/                 the steps + cc_writer.py (subscription routing), llm_client.py (timeouts,
                        fallback), deliver.py, operator_tz.py, bundle_floors.py, drop_reports.py,
                        thread_tracker.py ("What Changed"), critic.py, surface_render.py, ...
config/*.example.yaml   templates for models (all settings), sources, interests, WhatsApp chats
prompts/                analyst and writer prompts
sql/schema.sql          full schema (RLS on, service-role access only); sql/seed_*.sql starter data
examples/               a real sample brief
tests/                  run_all.py + one file per module

AI work uses the Claude plan's quota instead of a per-call bill. With the starter sources a run makes roughly 50 to 100 short claude -p calls (about 20 scoring batches of 10 articles, a similar number of classify batches, one write, a few summaries); a large source list needs a Max plan. The execution_log table records every step's model and tokens per run.

• 2026-10-01. Sources and interests as YAML files with manage.py sync; one-command database setup (manage.py db-init); a JSON copy of every brief and an optional post-save command for publishing; works on macOS and Linux; row-level security on every table; tests run on GitHub for Linux and Windows.
• 2026-09-29, first public release. Every AI step runs on a Claude subscription, with an alert whenever an API fallback is used; the brief is always saved as a file and phone delivery is optional; timezone set by name.

MIT (see LICENSE).