paperboy

Old paperboys delivered one newspaper to one house. This one reads ten thousand things and delivers the twelve that matter to you.

A daily Discord digest of news + research papers, scored by your local LLM against your own interests. No cloud, no SaaS, no account.

Drop your favorite RSS feeds in news_sources.yaml. Describe your stack in research-interests.md. Wake up to a Discord post like this:

🌅 morning prompts — today's queue:

📄 papers
• Diff-Transformer — applies to my-agent. Take?
• Self-Critique-via-Rubric vs. constitutional AI — applies to my-rag. Take?

📰 news
• Anthropic shipped tool-streaming. What changes for orchestrator+worker stacks?
• OpenAI's new pricing tier kills the gap for self-hosted Whisper. Hot take?

❓ answer
• How do you choose between Q3_K_M and Q4_K_M quantization for a 24GB GPU?

🎯 today
• Revive or kill: weekend-experiment (dead 17d) — pick one before EOD.

The bot also posts a separate Research Digest message with the top 3 papers scored ≥ 7/10 against your stack, each with a one-line "how WE'd apply it" note from your local LLM.

---

News → Agent Action Queue

A file-first bridge that promotes the highest-signal digest items (papers scored ≥ 7 or flagged should_prototype, plus the top-weighted prompts) into a reviewable queue of agent tasks — each with a concrete suggested_action.

It writes an append-only JSONL queue and a rendered Markdown view, and sends nothing (no Discord, no webhook). The queue is a human review gate: an agent acts only on items you've marked approved.

python -m paperboy.digest.action_queue          # write queue (JSONL + Markdown)
python -m paperboy.digest.action_queue --dry-run # preview only

If events.db is empty/absent, it falls back to a bundled public fixture so it always produces output.

• Module: paperboy/digest/action_queue.py
• Demo output: examples/action-queue.demo.jsonl · examples/action-queue.demo.md
• Dashboard: dashboard/action-queue.html (open directly in a browser)
• Walkthrough: docs/action-queue.md

---

What this gives you

Seven cron jobs that work together:

1. research-ingest — pulls new papers daily from arXiv + HF Daily Papers
   • Semantic Scholar.
2. research-score — local LLM rates each paper 0-10 for whether it could improve one of YOUR systems (grounded by research-interests.md).
3. news-opinion — fetches RSS feeds grouped by vertical, drafts a sharp hot-take prompt per item via local LLM.
4. papers-to-prompts — promotes top-scored papers into the morning queue.
5. topical-questions — picks AEO-shaped questions from YOUR pillar map.
6. today-briefing — parses a daily-brief markdown file (optional) for carry-overs and "dead repo" callouts.
7. prompt-digest — assembles + posts the morning digest to Discord.
8. research-digest — posts the top-3 newly-scored papers separately.

All of them write to one SQLite events.db. Everything is dedup'd by content hash, so re-runs are idempotent.

What you need

• A Linux box with systemd (any small VPS or homelab machine works)
• Ollama running locally with one fast model (e.g. llama3.2:3b) and one stronger model (e.g. qwen2.5:7b)
• A Discord webhook (or a bot token + channel ID)
• Python 3.10+

That's it. No Postgres, no Redis, no Docker, no cloud accounts.

Install

git clone https://github.com/cgallic/paperboy.git
cd paperboy
sudo ./scripts/bootstrap.sh

Then:

1. Edit /etc/paperboy/paperboy.env — set DISCORD_WEBHOOK.
2. Edit /etc/paperboy/news_sources.yaml — your RSS feeds, by topic.
3. Edit /etc/paperboy/research-interests.md — describe YOUR systems. This is the most important file — without it, the scorer can't tell what matters to you.
4. (Optional) Edit /etc/paperboy/topical-map.md — your content pillars.

Test:

sudo systemctl start paperboy-news-opinion.service
journalctl -u paperboy-news-opinion.service -n 50

See docs/INSTALL.md for the full walkthrough.

Architecture (one diagram)

            morning UTC                       

10:00 │ research-ingest  ──► events.db (paper events)
10:45 │ research-score   ──► events.db (paper-score events, via local Ollama)
11:00 │ news-opinion     ──► events.db (pattern-scan/question, stream=news)
11:05 │ papers-to-prompts ──► events.db (stream=papers)
11:08 │ topical-questions ──► events.db (stream=answer)
11:10 │ today-briefing   ──► events.db (stream=today)
11:15 │ prompt-digest    ──► Discord  (the morning post)
12:00 │ research-digest  ──► Discord  (separate top-3 paper post)

One SQLite table. Twelve env vars. Zero queues, zero brokers.

See docs/ARCHITECTURE.md for the full data model and event schema.

Customizing

Most of the value comes from tuning the config files, not the code. See docs/personalizing.md for:

• Adding/removing news verticals
• Writing a good research-interests.md (the scorer's grounding)
• Building your own topical map
• Optional: enabling LLM reframing of prompts into audience-targeted hooks

Relationship to open-brane

paperboy is a consumer of an event-log architecture defined by cgallic/open-brane — "an event-log brain for people who keep losing context." open-brane owns the events.db schema, the canonical write path (record_event.py), an MCP server, and adapters for gdrive / Claude sessions / git history. It also ships embed_events.py + semantic_search.py for vector queries.

paperboy adds the news + research-paper domain on top: RSS scanners, arXiv + HF + Semantic Scholar ingest, local-LLM scoring against your stack, and the morning Discord digest.

paperboy can run standalone — it ships a compatible schema in its own db.py so you don't need open-brane to get the digest working. But if you install open-brane alongside (or first), every other ingester populates the same store and paperboy's digest gets richer for free.

Roadmap

See docs/ROADMAP.md. The full planned stack:

1. open-brane — events.db foundation + MCP + canonical adapters (already shipped)
2. paperboy (this repo) — news + papers + Discord digest
3. More open-brane adapters — gmail, chatgpt, audible, linkedin, etc. (PRs to open-brane or sibling repos)
4. Markdown KG / wiki compiler — typed wiki pages + NL query CLI (planned, name TBD)
5. cgallic/pendant-android + cgallic/pendant-pipeline — custom Omi-pendant app + agent-side audio pipeline (planned)
6. cgallic/casa-a2a — JSON-RPC 2.0 broker + skill-sidecar pattern for cross-host agent coordination (planned)

License

MIT. See LICENSE.