Building News Agents for Daily News Recaps with MCP, Q, and tmux

Eugene Yan spent a couple of weekend hours building something most of us have wished for during doom-scrolling season: a system that reads Hacker News, TechCrunch, Wired, AI News, and two WSJ feeds so you don't have to. The tool, called news-agents, runs on Amazon Q CLI for the agentic brains and MCP (Model Context Protocol) for the actual news-fetching tools, with tmux handling the visual choreography of watching multiple agents work at once.

The architecture is refreshingly simple once you see it laid out. A main agent reads a feeds.txt file, splits six sources into three chunks, then spawns three sub-agents, each living in its own tmux pane. Each sub-agent chews through its assigned feeds, parses stories, tags them into categories like AI/ML or business, and writes a summary file. When all three report back, the main agent combines everything into a single markdown file. For May 4th, this pipeline processed 124 items across six sources and sorted them into 42 categories, with AI integration topics dominating at 25 percent of the total.

What makes this interesting isn't the AI summarization itself, which by 2025 standards is almost mundane. It's the plumbing. Each feed needs its own custom parser, because RSS formats vary wildly between Hacker News and something like WSJ's Dow Jones feed. Yan wrote a Python fetcher for the Hacker News RSS using httpx and ElementTree, then wrapped it in an MCP tool decorated with a simple @mcp.tool() call so the agent could invoke it directly. That's the whole trick of MCP: it turns a boring RSS parser into something an LLM can call as a tool, no fine-tuning or prompt gymnastics required.

Yan admits the setup has rough edges. He initially wanted to deploy this as a hosted web app, maybe on something like Daytona, but building remote MCP servers turned out to be more work than a couple of free hours could handle. So for now it's a local CLI tool you run yourself, cloning the repo, syncing dependencies with uv, and firing up Q chat with a context file that tells the main agent to read its instructions and spin up sub-agents. He's already eyeing other use cases, like parsing internal design docs or running multi-agent coding workflows, which suggests this pattern of split-work-then-recombine is going to show up in a lot more places than just morning news digests.