agents-radar

English | 中文

A GitHub Actions workflow that runs every morning at 07:00 CST. It aggregates AI ecosystem signals from 10 data sources, then publishes bilingual (Chinese + English) daily digests as GitHub Issues and committed Markdown files.

Data Sources

Web UI

https://duanyytop.github.io/agents-radar

Browse all historical digests in a clean, dark-themed interface — no login required. Reports are rendered from the Markdown files in this repo via GitHub Pages.

Telegram Channel & Feishu Group

Subscribe to get daily digest notifications pushed directly to your preferred platform. Each message links to all reports for that day (ZH and EN variants) plus the Web UI and RSS feed.

RSS Feed

https://duanyytop.github.io/agents-radar/feed.xml

Subscribe in any RSS reader (Feedly, Reeder, NewsBlur, etc.) to receive new digests automatically. The feed includes the latest 30 reports across all report types, updated daily alongside manifest.json.

MCP Server

https://agents-radar-mcp.duanyytop.workers.dev

A hosted Model Context Protocol server that exposes agents-radar data as tools. Any MCP-compatible client (Claude Desktop, OpenClaw, etc.) can query the latest AI ecosystem reports directly.

Available tools:

Claude Desktop setup — add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "agents-radar": {
      "url": "https://agents-radar-mcp.duanyytop.workers.dev"
    }
  }
}

Restart Claude Desktop after saving. You can then ask Claude things like:

• "What's the latest in AI CLI tools?" → calls get_latest
• "Search for Claude Code mentions this week" → calls search
• "Show me the AI trending report for 2026-03-05" → calls get_report

OpenClaw setup — run the following command:

openclaw mcp add --transport http agents-radar https://agents-radar-mcp.duanyytop.workers.dev

Or add it manually to ~/.openclaw/openclaw.json:

{
  "mcpServers": {
    "agents-radar": {
      "type": "http",
      "url": "https://agents-radar-mcp.duanyytop.workers.dev"
    }
  }
}

You can then ask OpenClaw things like:

• "What's the latest in AI CLI tools?" → calls get_latest
• "Search for Claude Code mentions this week" → calls search
• "Show me the AI trending report for 2026-03-05" → calls get_report

Self-hosting — deploy your own instance from the mcp/ directory:

cd mcp
pnpm install
wrangler deploy

Tracked sources

AI CLI tools (GitHub)

Repos marked discussions: true in config.yml (Codex, Pi) also have their GitHub Discussions pulled in.

Claude Code Skills (GitHub)

PRs and issues are fetched without a date filter and sorted by popularity (comment count), so the report always reflects the most actively discussed skills — not just the newest.

OpenClaw + AI agent ecosystem (GitHub)

OpenClaw is tracked as the primary reference project, alongside several peer projects in the personal AI assistant / autonomous agent space for cross-ecosystem comparison.

AI infrastructure (GitHub)

The inference and serving layer the agent/CLI tools run on top of — tracked in a dedicated report with its own cross-project comparison.

Summaries focus on new model/hardware support, performance work, breaking changes, and what each change means for developers building on top of these projects.

GitHub AI Trending

Two data sources are fetched in parallel every day:

The LLM filters out non-AI repos from the trending list, classifies the rest by dimension (AI infrastructure / agents / applications / models / RAG), and extracts trend signals.

Hacker News

Top AI stories from the last 24 hours, fetched via the Algolia HN Search API. Six queries run in parallel (AI, LLM, Claude, OpenAI, Anthropic, machine learning), results are deduplicated and ranked by points. The top 30 stories are passed to the LLM for analysis.

Official web content (sitemap-based)

New articles are detected by comparing sitemap lastmod timestamps against a persisted state file (digests/web-state.json). On the first run, up to 25 recent articles per site are fetched and a comprehensive overview report is generated. On subsequent runs, only new or updated URLs trigger a report; if nothing changed, the web report step is skipped entirely.

Features

• Fetches issues, pull requests, and releases updated in the last 24 hours across all tracked repos
• Tracks trending Claude Code Skills — sorted by community engagement, not recency
• Generates a per-tool summary for each CLI repository and a cross-tool comparative analysis
• Generates a deep OpenClaw project report plus a cross-ecosystem comparison against 4 peer projects
• Tracks 6 AI infrastructure projects (inference engines, gateways, fine-tuning) with a dedicated report and cross-project comparison
• Scrapes official Anthropic and OpenAI web content via sitemaps; detects new articles incrementally
• Monitors GitHub Trending daily + searches 6 AI topic tags; classifies repos by dimension and extracts trend signals
• Fetches top-30 AI stories from Hacker News (last 24h, ranked by points); generates community sentiment report
• Publishes GitHub Issues for each report type; commits Markdown files to digests/YYYY-MM-DD/
• Generates every report body once in English and translates it to Chinese, instead of running the whole pipeline twice per language
• Runs on a daily schedule via GitHub Actions; supports manual triggering
• All tracked repositories are configurable via config.yml — no code changes needed

Setup

1. Fork this repository

2. Customize config.yml (optional)

Edit config.yml in the repo root to add, remove, or replace the tracked repositories. The file is fully commented. No code changes are needed — the pipeline reads it on every run and falls back to built-in defaults if the file is absent.

# Add a new CLI tool
cli_repos:
  - id: my-tool
    repo: owner/my-ai-cli
    name: My AI Tool

# Add a new peer project to the OpenClaw ecosystem comparison
openclaw_peers:
  - id: my-agent
    repo: owner/my-agent
    name: My Agent

# Add a new AI infrastructure project
infra_repos:
  - id: my-engine
    repo: owner/my-engine
    name: My Engine
    paginated: true   # for repos with >100 daily issue/PR updates

3. Add Secrets

Go to Settings → Secrets and variables → Actions and add:

GITHUB_TOKEN is provided automatically by GitHub Actions. When using github-copilot as the provider, the same GITHUB_TOKEN is used for LLM calls.

Setting up Telegram notifications (optional):

1. Message @BotFather on Telegram, create a bot, and copy the token
2. Add the bot to your channel/group, or start a DM with it
3. Get the chat ID via @userinfobot or the getUpdates API
4. Add TELEGRAM_BOT_TOKEN and TELEGRAM_CHAT_ID as repository secrets

If neither secret is set, the notification step is silently skipped.

3. Enable the workflow

Confirm the workflow is enabled in the Actions tab.

To test immediately, go to Actions → Daily Agents Radar → Run workflow.

First run note: The web content step will fetch up to 50 articles (25 per site) and may take a few extra minutes. Subsequent runs are fast — only new articles are processed.

LLM providers

Set LLM_PROVIDER to choose which model backend powers the digest generation. Defaults to anthropic.

Override the model name with ANTHROPIC_MODEL, OPENAI_MODEL, GITHUB_COPILOT_MODEL, OPENROUTER_MODEL, DEEPSEEK_MODEL, or QWEN_MODEL respectively. The Qwen endpoint can be overridden with DASHSCOPE_BASE_URL.

The scheduled daily run uses qwen / qwen-flash.

The provider abstraction lives in src/providers/ — each provider is a separate file implementing the LlmProvider interface. Adding a new provider only requires creating a new file and registering it in the factory.

Running locally

pnpm install

export GITHUB_TOKEN=ghp_xxxxx

# Option A: Anthropic (default)
export ANTHROPIC_API_KEY=sk-ant-xxxxxxxx

# Option B: OpenAI
# export LLM_PROVIDER=openai
# export OPENAI_API_KEY=sk-xxxxxxxx

# Option C: GitHub Copilot (uses GITHUB_TOKEN)
# export LLM_PROVIDER=github-copilot

# Option D: OpenRouter
# export LLM_PROVIDER=openrouter
# export OPENROUTER_API_KEY=sk-or-xxxxxxxx

# Option E: DeepSeek
# export LLM_PROVIDER=deepseek
# export DEEPSEEK_API_KEY=sk-xxxxxxxx

# Qwen (Alibaba Model Studio)
# export LLM_PROVIDER=qwen
# export DASHSCOPE_API_KEY=sk-xxxxxxxx

export DIGEST_REPO=your-username/agents-radar  # optional; omit to only write files

pnpm start

Output format

Files are written to digests/YYYY-MM-DD/:

A shared state file digests/web-state.json tracks which web URLs have been seen; it is committed alongside the daily digests.

Each report is generated in both Chinese (ai-cli.md) and English (ai-cli-en.md). The Web UI sidebar shows ZH / EN toggle buttons for reports that have both variants.

---

ai-cli.md / ai-cli-en.md structure:

## Cross-Tool Comparison
  Ecosystem overview / Activity comparison table / Shared themes / Differentiation / Trend signals

## Per-Tool Reports
  <details> Claude Code    — [Claude Code Skills Highlights]
                             Top skills / Community demand trends / High-potential pending skills
                             ---
                             Today's summary / Hot issues / PR progress / Trends
  <details> OpenAI Codex   — Today's summary / Hot issues / PR progress / Trends
  <details> Gemini CLI     — ...
  <details> GitHub Copilot CLI — ...
  <details> OpenCode       — ...
  <details> Pi             — ...
  <details> Qwen Code      — ...

ai-agents.md / ai-agents-en.md structure:

Issues: N | PRs: N | Projects covered: 5

## OpenClaw Deep Dive
  Today's summary / Releases / Project progress / Community highlights /
  Bug stability / Feature requests / User feedback / Backlog

## Cross-Ecosystem Comparison
  Ecosystem overview / Activity table / OpenClaw positioning /
  Shared technical directions / Differentiation / Community maturity / Trend signals

## Peer Project Reports
  <details> ZeroClaw     — Today's summary / Releases / Progress / ... (8 sections)
  <details> Hermes Agent — ...
  <details> IronClaw     — ...
  <details> QwenPaw      — ...

ai-infra.md / ai-infra-en.md structure:

Projects covered: 6

## Cross-Project Comparison
  Ecosystem overview / Activity table / Model support race /
  Performance frontier / Layer positioning / Trend signals

## Per-Project Reports
  <details> vLLM       — Today's highlights / Releases & breaking changes /
                         New model & hardware support / Performance & optimization /
                         Stability & regressions / What it means for app developers
  <details> SGLang     — ...
  <details> llama.cpp  — ...
  <details> Ollama     — ...
  <details> LiteLLM    — ...
  <details> Unsloth    — ...

ai-web.md / ai-web-en.md structure:

Sources: anthropic.com (N articles) + openai.com (N articles)

Today's summary
Anthropic / Claude highlights  (news / research / engineering / learn)
OpenAI highlights              (research / release / company / safety / ...)
Strategic signals
Notable details
[First full crawl also includes: Content landscape overview]

ai-trending.md / ai-trending-en.md structure:

Sources: GitHub Trending + GitHub Search API

Today's summary
Top repos by dimension
  🔧 AI Infrastructure  — frameworks / SDKs / inference engines / CLIs
  🤖 AI Agents          — agent frameworks / multi-agent / automation
  📦 AI Applications    — vertical products / solutions
  🧠 Models & Training  — model weights / training frameworks / fine-tuning
  🔍 RAG & Knowledge    — vector databases / retrieval augmentation
Trend signal analysis
Community focus

ai-hn.md / ai-hn-en.md structure:

Sources: Hacker News (top-30 AI stories, last 24h)

Today's summary
Top stories & discussions
  🔬 Models & Research  — new model releases / papers / benchmarks
  🛠️ Tools & Engineering — open-source projects / frameworks / engineering practice
  🏢 Industry news      — company news / funding / product launches
  💬 Opinions & debate  — Ask HN / Show HN / hot threads
Community sentiment signals
Worth reading

Historical digests are stored in digests/. Published issues are tagged by type: digest · openclaw · web · trending · hn · ph · arxiv · hf · community.

Only the most recent day's issues are left open — pnpm close-stale closes everything older on each run, so the issue list stays at roughly one day's worth instead of thousands. Nothing is deleted: the label links above include closed issues, and the full archive also lives under digests/ and in the Web UI. Issues without a digest label, and pull requests, are never touched.

Weekly and monthly rollup reports were discontinued in July 2026; past ones remain browsable under digests/ and in the Web UI.

Schedule

To change the schedule, edit the cron expression in .github/workflows/daily-digest.yml.

GitHub's scheduled runs are queued, not guaranteed — typical delay for this workflow is 10-15 minutes, so a 06:37 CST start lands the digest around 07:00 CST. The minute is off the hour on purpose: the :00 slot is the most contended, and while this workflow sat at 0 23 * * * the delays degraded from minutes to hours (the 2026-08-26 run was dispatched 5h07m late, and the 2026-08-27 run was never created at all).

When a scheduled run is late enough that you dispatch a catch-up run manually, the two would otherwise both build the same day's digest and open a duplicate set of issues. Two guards prevent that: a concurrency: daily-digest group serializes overlapping runs, and a guard job skips any scheduled run whose digests/YYYY-MM-DD folder (CST date) is already committed. Manual workflow_dispatch runs always proceed, so you can still force a regeneration.

Star History