Herald

Local-first news intelligence for AI agents

claude plugin marketplace add heurema/emporium
claude plugin install herald@emporium

What it does

Herald collects articles from RSS feeds and Hacker News, deduplicates by URL, clusters related articles into stories using title similarity, scores them by source weight and recency, and generates a ranked Markdown digest — all locally, no API keys, no cloud.

The pipeline: collect → ingest → cluster → project.

RSS/Atom feeds ─┐
                 ├─→ articles ─→ stories (clustered) ─→ scored brief
HN Algolia API ─┘

Install

claude plugin marketplace add heurema/emporium
claude plugin install herald@emporium

Manual install from source

git clone https://github.com/heurema/herald
cd herald
pip install httpx fastfeedparser pyyaml

Then symlink or copy the plugin directory into Claude Code's plugin path and run /news-init.

Quick start

/news-init          # creates ~/.herald/ with config and database
/news-add <url>     # add an RSS feed
/news-run           # collect, ingest, cluster, generate brief
/news-digest        # read the latest brief

Commands

Command	What it does
/news-init	Create data directory, config template, and database
/news-add <url>	Add an RSS/Atom feed — auto-discovers feed URL
/news-sources	View all sources grouped by category
/news-sources remove <name>	Remove a source from config
/news-run	Run the full pipeline manually
/news-digest	Read the latest brief with 5-section analysis
/news-status	Show article/story counts and last run time
/news-stop	Show cleanup options

Architecture

Herald v2 uses a 4-stage pipeline with SQLite storage:

1. Collect — fetches RSS/Atom feeds and HN front-page stories via public APIs. Optional Tavily adapter for web search.
2. Ingest — UPSERT articles with URL canonicalization, deduplication, topic assignment, and cross-source mention tracking.
3. Cluster — groups related articles into stories using SequenceMatcher title similarity with 4 merge guards (threshold, time gap, title length, version/number conflict). Canonical article re-election with hysteresis.
4. Project — generates a Markdown brief with YAML frontmatter, stories grouped by type (release, research, tutorial, opinion, news), scored and ranked.

Data model

sources → articles → mentions (cross-source)
                  → article_topics
                  → story_articles → stories → story_topics

Scoring

• Article score: source_weight + min(points/500, 3.0) + keyword_density * 0.2 + release_boost
• Story score: max(article_scores) + log(source_count) * 0.3 + momentum

Configuration

Config: ~/.herald/config.yaml

sources:
  - id: hn
    name: Hacker News
    type: hn           # hn | rss | tavily
    weight: 0.3
    category: community
  - id: simonw
    name: Simon Willison
    type: rss
    url: https://simonwillison.net/atom/everything/
    weight: 0.25
    category: community

topics:
  ai_agents:
    keywords: [agent, agents, agentic, tool use, mcp]
  ai_models:
    keywords: [claude, gpt, gemini, llama, llm]

clustering:
  threshold: 0.65        # title similarity threshold
  max_time_gap_days: 7   # max days between clustered articles

schedule:
  interval_hours: 4

Data paths

~/.herald/
├── config.yaml
├── herald.db           # SQLite database
└── briefs/
    └── {run_id}.md     # generated briefs

Requirements

• Python 3.12+
• httpx, fastfeedparser, pyyaml
• macOS or Linux
• Claude Code

Privacy

Herald fetches only public RSS feeds and the HN Algolia API. All data stays on your machine under ~/.herald/. No telemetry, no cloud sync. Optional Tavily adapter requires a free API key but is not needed for core functionality.

Feedback

Found a bug? All heurema plugins ship with Reporter - file issues without leaving Claude Code:

claude plugin install reporter@emporium
/report bug

See also

• Herald v2: Local-First News Intelligence for AI Agents - blog post with architecture deep-dive
• emporium - plugin marketplace
• signum - contract-first AI dev pipeline
• reporter - issue filing from Claude Code

License

MIT