🚀 femtech-radar

FemTech intelligence as a GitHub-native, agent-driven pipeline

A project of FemTech Weekend — the live site shares the org's brand identity.

Agent-first FemTech intelligence: an MCP server that fetches, dedupes, and scores women's-health & FemTech research and industry news, designed to feed GitHub Agentic Workflows and an auto-updating Astro + RSS site.

Live Demo · Documentation · Changelog · Report Bug · Request Feature

• 🌟 Introduction

• ✨ Key Features

• 🛠️ Tech Stack

• 🏗️ Architecture

• 🚀 Getting Started

• 🛳 Project Status & Roadmap

• 📖 Usage Guide

• ⌨️ Development

• 🤝 Contributing

• ❤️ Sponsor

• 📄 License

• 🙋‍♀️ Author

🌟 Introduction

The FemTech and women's-health world produces a scattered firehose of signal — research preprints, funding and product news, community opportunities, technical discussion — across dozens of sources. femtech-radar turns that noise into a curated, deduplicated, ranked digest, using only GitHub-native primitives plus a reusable Model Context Protocol (MCP) server.

It's a complete, live three-unit pipeline — scrape → curate → publish:

1. an MCP server (the deterministic "brain") fetches from multiple sources, normalizes into one shape, dedupes, and scores each item by relevance × popularity × freshness, exposing the result over MCP;

2. a weekly GitHub Agentic Workflow (gh aw) drives the MCP, lets Copilot curate the digest, and commits the week's data/*.json via a review-gated PR;

3. an Astro + RSS site reads that data and auto-deploys to GitHub Pages.

It's built for FemTech / women-in-tech practitioners who want signal without the noise — and as a reference implementation of the gh aw × MCP × GitHub Pages pattern. The product is live: a subscribable weekly intelligence site at https://chanmeng666.github.io/femtech-radar/ that updates itself for free on a public GitHub repo.

✨ Key Features

1 One MCP tool, a whole radar — radar_collect returns normalized, deduped, scored items per section. All five sections are live: industry (Google News), research (arXiv), opportunities (LinkedIn, with opt-in SerpAPI Google Jobs), discussions (Hacker News + Mastodon), and china — a China-focused "China Watch" lens on Chinese women's-health / FemTech + 投融资 (funding) news (Google News 中文 + 36氪 + 新浪财经) plus China-scoped research (PubMed + ClinicalTrials.gov).

2 Deterministic, testable core — fetch → normalize → dedupe → score. The server makes no editorial judgment; that's left to the agent that drives it. 46 unit tests, and zero real network calls in tests (all I/O is injected).

3 Pluggable source adapters — each source is one file behind a uniform Adapter interface. Add a feed by writing a collect() that returns RadarItem[].

4 Resilient by construction — every source failure degrades to an empty result plus a warning; a malformed URL or date never throws out of a run.

5 GitHub-native, near-zero cost — designed to run inside GitHub Actions on a public repo, where Actions and Pages are free; the only metered resource is a few Copilot credits per week.

6 Reusable anywhere MCP runs — drop it into Claude Desktop or any MCP client; it isn't coupled to this project.

7 Live, self-updating, subscribable — a weekly gh aw workflow curates the digest into a review-gated data/*.json PR, and an Astro + RSS site rebuilds itself on GitHub Pages. Readers subscribe via the full RSS feed or per-section feeds at /rss/<section>.xml — no manual scraping, no server to run.

🛠️ Tech Stack

• Language / Runtime: TypeScript 5 (strict, ESM) · Node.js ≥ 20

• MCP: @modelcontextprotocol/sdk (stdio server)

• Parsing / validation: Zod (schema & runtime validation) · fast-xml-parser (Atom/RSS)

• Tooling: pnpm workspaces (monorepo) · Vitest (tests) · tsup (build)

• Orchestration: GitHub Agentic Workflows (gh aw) · engine copilot, model gpt-4.1 · weekly schedule → review-gated data PR

• Site: Astro 5 (femtech-radar-site) · @astrojs/rss · GitHub Pages deploy via deploy-pages.yml

• Design: vanilla CSS design system (site/src/styles/global.css) aligned to the FemTech Weekend brand — warm-brown editorial palette, Georgia serif + system sans, sharp corners, light mode only. Full reference: docs/design-system.md

• Built on: the opportunities adapter ports LinkedIn guest-endpoint logic from the owner's linkedin-jobs-search project (inspired by linkedin-jobs-api); the opt-in SerpAPI Google Jobs path is ported from server-google-jobs

🏗️ Architecture

graph TD
    subgraph Sources
      A1[arXiv API]
      A2[Google News RSS<br/>+ 中文]
      A3[LinkedIn Jobs]
      A4[Hacker News + Mastodon]
      A5[36氪 · 新浪财经]
      A6[PubMed · ClinicalTrials.gov]
    end
    A1 --> AD[Source adapters<br/>normalize → RadarItem]
    A2 --> AD
    A3 --> AD
    A4 --> AD
    A5 --> AD
    A6 --> AD
    AD --> P["collect()<br/>dedupe → score → sort → since-filter"]
    P --> T[MCP tools<br/>radar_collect · radar_sources]
    T --> C[MCP clients]
    C --> W[Weekly gh aw workflow<br/>Copilot curates · writes why_it_matters]
    W --> D[(data/YYYY-Www.json<br/>review-gated PR)]
    D --> S[Astro + RSS site<br/>GitHub Pages · live]

All three units are built and live; all five source sections are active.

Separation of determinism vs judgment: the MCP server does only deterministic work (fetch, normalize, dedupe, score). Editorial choices — which items to feature, how to summarize — belong to the agent (gh aw + Copilot) that drives it. This keeps the core unit-testable and reusable.

🚀 Getting Started

Prerequisites

• Node.js ≥ 20 (the server uses global fetch)

• pnpm ≥ 9 (npm i -g pnpm)

• (optional) an MCP client such as Claude Desktop, or the gh aw CLI

Installation

git clone https://github.com/ChanMeng666/femtech-radar.git
cd femtech-radar
pnpm install

# Build the MCP server
pnpm --filter @chanmeng666/femtech-radar-mcp build

# Run the test suite (46 tests)
pnpm --filter @chanmeng666/femtech-radar-mcp test

The built server is an executable stdio MCP server at packages/mcp-server/dist/index.js. It's also published to npm, so any MCP client can run it with npx -y @chanmeng666/femtech-radar-mcp.

Run the site locally

# Dev server (reads data/*.json)
pnpm --filter femtech-radar-site dev

# Production build + preview
pnpm --filter femtech-radar-site build
pnpm --filter femtech-radar-site preview

The published site is live at https://chanmeng666.github.io/femtech-radar/ (RSS: /rss.xml); it rebuilds automatically whenever a new weekly data/*.json lands on master.

🛳 Project Status & Roadmap

All three planned layers are built and live in production (see docs/superpowers/specs for the full design and docs/superpowers/plans for the per-unit implementation plans):

• ✅ v1 — MCP server: industry (Google News) + research (arXiv) adapters, dedupe/score pipeline, radar_collect / radar_sources tools, resilient error handling, 46 tests. Published to npm as @chanmeng666/femtech-radar-mcp.

• ✅ v2 — orchestration: a weekly gh aw workflow (engine copilot, model gpt-4.1) that drives the MCP, curates a digest, and emits a review-gated data PR plus a summary issue — proven end-to-end in production (first digest: data/2026-W27.json).

• ✅ v3 — publishing (live): Astro 5 site auto-deployed to GitHub Pages at https://chanmeng666.github.io/femtech-radar/ with a subscribable RSS feed at https://chanmeng666.github.io/femtech-radar/rss.xml; rebuilds automatically on every weekly data update.

• ✅ vNext — 4-section pipeline: opportunities (LinkedIn, with opt-in SerpAPI Google Jobs) and discussions (Hacker News + Mastodon) adapters now active; per-section RSS feeds at /rss/<section>.xml; site polish (favicon, numeric-entity decode, sources chip); weekly workflow updated to collect all four sections.

• ✅ China Watch — 5th section (npm 0.4.1): a China-focused china section merging Chinese women's-health / 投融资 news (Google News 中文 + 36氪 + 新浪财经) and China-scoped research (PubMed + ClinicalTrials.gov); Chinese titles with English why_it_matters. Live on the site as 05 · CHINA with a /rss/china.xml feed. Includes a CJK-aware dedupe fix and a women's-health-specific filter for the general Chinese feeds. Structured funding-round data (paid databases) is intentionally out of scope. See docs/operations.md.

Deferred to a future version: ChatOps slash commands (/deep-dive), full bilingual support (i18n routing is reserved), a newsletter, femtech-weekend.com domain integration, and a structured funding-round database.

📖 Usage Guide

femtech-radar is consumed as an MCP server. It exposes two tools:

All five sections are live. industry (Google News) and research (arXiv) require no API key. opportunities uses LinkedIn by default (opt-in SerpAPI Google Jobs when SERP_API_KEY is set). discussions merges Hacker News Algolia + Mastodon hashtag timelines (both free, no key). china merges Google News 中文 + 36氪 + 新浪财经 (Chinese women's-health / 投融资 news, keyword-filtered) + PubMed + ClinicalTrials.gov (China-scoped research) — all free, no key. Every source degrades gracefully to [] if unavailable.

Subscribe via RSS: the weekly digest is published at https://chanmeng666.github.io/femtech-radar/rss.xml and works in any feed reader.

Use with GitHub Agentic Workflows (gh aw)

mcp-servers:
  femtech-radar:
    command: npx
    args: ["-y", "@chanmeng666/femtech-radar-mcp"]

Use with Claude Desktop

{
  "mcpServers": {
    "femtech-radar": {
      "command": "node",
      "args": ["/absolute/path/to/femtech-radar/packages/mcp-server/dist/index.js"]
    }
  }
}

The package is published to npm, so npx -y @chanmeng666/femtech-radar-mcp works out of the box (current version 0.4.1). The local dist/index.js path above is an alternative for development.

See packages/mcp-server/README.md for the full tool reference.

⌨️ Development

pnpm install                                              # install workspace deps
pnpm --filter @chanmeng666/femtech-radar-mcp test         # run tests (Vitest)
pnpm --filter @chanmeng666/femtech-radar-mcp build        # build with tsup

Project layout

packages/mcp-server/src/
├── schema.ts          # Zod RadarItem / WeeklyData (the shared data contract)
├── dedup.ts           # URL canonicalization + title-similarity dedupe
├── score.ts           # relevance × popularity × freshness scoring
├── adapters/          # one file per section (industry = Google News, research = arXiv, opportunities = LinkedIn/SerpAPI, discussions = HN + Mastodon, china = Google News 中文 + 36氪 + Sina + PubMed + ClinicalTrials)
├── collect.ts         # orchestration: adapter → dedupe → score → sort → since-filter
├── tools.ts           # radar_collect / radar_sources handlers
└── index.ts           # stdio MCP server entry

Adding a source adapter: implement the Adapter interface in adapters/, return RadarItem[] from collect(opts) using the injected fetcher (never call fetch directly — that keeps it testable), then wire it into ADAPTERS in collect.ts.

See AGENTS.md for AI-agent-oriented project conventions and gotchas, and docs/operations.md for the operational runbook (weekly workflow, Copilot credits/quota, engine choices, npm publish, and manual data seeding).

🤝 Contributing

Contributions make the open-source community an amazing place to learn and create. Please read the Contributing Guide and the Code of Conduct before you start, and use the provided issue / pull-request templates.

❤️ Sponsor

If this project helps you, please consider supporting its development:

For questions and help, see SUPPORT.md. For security issues, see SECURITY.md.

📄 License

This project is released under the MIT license.

🙋‍♀️ Author

Chan Meng