agents-radar: A GitHub Actions Pipeline for Daily AI Ecosystem Digests

What agents-radar Collects and Who Uses It

agents-radar is a tool for engineers and researchers who want a structured daily summary of what is moving in the AI ecosystem without building their own data pipeline. The README describes a GitHub Actions workflow that fires every morning at 07:00 CST and reads from 10 data sources: GitHub repository activity, GitHub Trending, Hacker News, ArXiv preprints, Hugging Face models, Product Hunt, Dev.to, Lobste.rs, and the official sitemaps of Anthropic and OpenAI.

The workflow writes bilingual digests in both Chinese and English. Each digest is committed as a Markdown file under the digests/ directory and also published as a GitHub Issue. A web interface at https://duanyytop.github.io/agents-radar lets anyone browse the archive without a login.

The tool is focused on AI tooling rather than the entire ML research field. The README's tracked repository list covers AI coding assistants (Claude Code, OpenAI Codex, Gemini CLI, GitHub Copilot CLI) and agent platforms (OpenClaw, Hermes Agent), not foundational model research. Teams monitoring a different slice of the AI space will find the hardcoded source list in config.yml either a good fit or a reason to fork.

The Data Pipeline: Sources, Scheduling, and Report Generation

Each data source is handled separately. GitHub API calls pull issues, pull requests, and releases from 18 tracked repositories. Hacker News queries run six parallel requests via the Algolia API, retrieving the top 30 AI stories from the previous 24 hours. ArXiv returns the latest papers from three categories: cs.AI, cs.CL, and cs.LG, within a 48-hour window. Hugging Face trending models are fetched only on Mondays. The README labels this explicitly as a weekly update, meaning a model that launches on a Tuesday will not appear until the following Monday's digest.

The Anthropic and OpenAI sources use sitemap comparison. The workflow diffs the lastmod timestamp in each sitemap against the previous snapshot to identify newly published articles, so it reports additions rather than the full sitemap.

An LLM writes the summaries. The .env.example file documents five supported providers: Anthropic (the default, accessed via ANTHROPIC_API_KEY), OpenAI, GitHub Copilot, OpenRouter, and DeepSeek. The provider is selected via the LLM_PROVIDER environment variable. The package.json confirms that both @anthropic-ai/sdk and the openai package are runtime dependencies, so both Anthropic and OpenAI providers are available without additional installs.

Running agents-radar Locally

Clone the repository and install dependencies with pnpm:

pnpm install

Copy the environment template and fill in the values you need:

cp .env.example .env

At minimum, GITHUB_TOKEN and an LLM API key are required. For the default Anthropic provider, the .env.example shows these two entries:

GITHUB_TOKEN=ghp_xxxxx
ANTHROPIC_API_KEY=sk-ant-xxxxx

The .env.example notes that the file is not loaded automatically. Before running any script, export the variables in your shell:

set -a; source .env; set +a

Then start the pipeline:

pnpm start

This runs tsx src/index.ts, which executes the full aggregation and summary pipeline. The result is a bilingual digest written to digests/. If DIGEST_REPO is set in .env (commented out by default), the workflow also publishes the digest as a GitHub Issue to that repository. In GitHub Actions, the environment variables are provided as repository secrets rather than a .env file.

MCP Server: Querying Digests from Claude Desktop or OpenClaw

A hosted MCP server at https://agents-radar-mcp.duanyytop.workers.dev exposes the digest archive as four tools: list_reports (lists available dates and report types), get_latest (fetches the most recent report of a given type), get_report (fetches a specific report by date and type), and search (keyword search across recent reports).

To connect Claude Desktop, add the following to ~/Library/Application Support/Claude/claude_desktop_config.json and restart the application:

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

For OpenClaw, the README gives a one-line command:

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

To self-host the MCP server on Cloudflare Workers, deploy from the mcp/ subdirectory:

cd mcp
pnpm install
wrangler deploy

The MCP integration makes agents-radar reports queryable from inside a conversation without opening a browser tab. The README gives three example queries: asking about the latest in AI CLI tools, searching for mentions of a specific project, and retrieving a report by date.

What agents-radar Does Not Cover

The source list is fixed in config.yml. Adding a new data source requires writing a fetcher module and registering it in the configuration. The README does not document this extension process, so a developer adding a new source must read the existing fetchers in src/ to understand the expected interface.

Hugging Face model data runs on a weekly schedule, not daily. The README states this explicitly: the Hugging Face source is tagged as weekly, Mondays only. Teams that need daily model tracking would have to change this cadence themselves.

The digests are generated by an LLM without human review. Factual errors in summaries are possible, particularly for source material where the LLM has limited context. The README does not document a correction or retraction workflow.

The tool covers cs.AI, cs.CL, and cs.LG on ArXiv. Topics like computer vision, robotics, or tabular data fall outside the tracked categories unless a specific paper happens to be cross-listed. Users monitoring a broader slice of ML research will need separate tooling for those domains.

Comparison with RSS Aggregators, Maintenance, and License

RSS aggregators like Feedly or Inoreader let a user subscribe to any AI-related feeds and read them in a single interface. The difference is that they present raw items from whatever feeds the user subscribes to; they do not cross-reference sources, generate summaries, or produce a unified narrative. agents-radar consolidates ten sources with a fixed set of AI tooling signals and writes a digest each morning rather than presenting a link list.

A human-curated AI newsletter is the other common alternative. The distinction is cadence and curation method: a newsletter editor selects stories by judgment; agents-radar selects by API query and threshold rules, then writes summaries via LLM. The automation means agents-radar updates every morning without editorial lag, but it cannot judge story significance the way a human editor can.

The last push to the master branch was on 2026-09-25. The repository is not archived.

The project is released under the MIT license, which permits commercial use, modification, and redistribution with the original copyright notice retained. Because the runtime calls third-party APIs (GitHub, ArXiv, Hugging Face, and others), users running agents-radar against those APIs are subject to each provider's terms of service independently of the MIT license. The LLM calls in particular will incur API costs depending on which provider is configured.

Editorial conclusion

Engineers who need a zero-setup, daily briefing on AI CLI tools and agent platforms will find agents-radar immediately usable after filling in a GitHub token and an LLM key in .env and running pnpm start. The fixed source list in config.yml makes it a poor fit for monitoring domains outside AI tooling. Teams that need human editorial judgment about story significance, or weekly rather than daily updates for most of the corpus, will hit the tool's structural limits quickly. Before forking to add a new data source, read the fetcher modules in src/ to confirm that a common interface exists for new sources.

Frequently asked questions

Official sources

1. Official documentation
2. Official README
3. Project repository