Open Source Radar — August 31: what agents read, teach, and plug into
The daily board is full of things that bolt capability onto agents you already run: a research skill that swaps search-engine editors for crowd votes, a classroom that builds courses with an agent instead of a prompt box, and the two plumbing layers — MCP servers and realtime voice — that keep showing up underneath everything else.
last30days-skill (Go, ~60,600 stars) — An agent skill that searches Reddit, X, YouTube, Hacker News, Polymarket, TikTok and the open web in parallel, then scores what it finds by upvotes, likes, and the money actually staked in prediction markets before an agent judge writes the brief. The pitch is pointed: Google aggregates editors, this searches people. Reddit, HN, Polymarket and GitHub work with zero configuration; a setup wizard unlocks X, YouTube, TikTok and arXiv once you bring your own keys. You'd reach for it when you need the last month of real sentiment on something — a model release, a framework, a market — and don't trust a search index to surface what practitioners actually said.
OpenMAIC (TypeScript, ~22,300 stars) — Tsinghua's multi-agent classroom turns one prompt into a full immersive course: slides, quizzes, interactives, narration, all generated and laid out as a live lesson rather than a document dump. Version 1.0.0 landed on August 27 with a chat-first agent workbench that plans a curriculum, builds and revises every page, and pulls from documents, audio, video or web search that you upload. Sessions are server-backed, so a long generation survives restarts and you can cancel, resume, and steer mid-run. It's deliberately provider-neutral — bring your own models, media, search and storage — which makes it a plausible open alternative to hosted course builders for training teams and teachers.
awesome-mcp-servers (~93,500 stars) — The canonical index of Model Context Protocol servers is back on the daily board and closing on a hundred thousand stars, which says more about MCP's grip than about any single entry. It aggregates the reference implementations and community servers across categories — data, cloud, dev tools, search, memory — so you can find the connector you need instead of writing another one. Treat it as a directory, not a vetted store: quality varies wildly and nothing here has passed a security review, which is exactly why it's still worth reading before you wire an agent to your production database.
LiveKit Agents (Python, nearly 14,000 stars) — The framework for realtime voice and video agents that run as server-side participants in a WebRTC room, now back on trending as voice moves from demo to deployment. You mix and match your own speech-to-text, LLM, text-to-speech or realtime API, it handles job scheduling and dispatch across users, and telephony integration lets an agent pick up or place actual phone calls. Semantic turn detection — a transformer that decides when you've finished talking — is the quiet feature that makes conversations stop feeling like walkie-talkie traffic. Apache-2.0 and self-hostable end to end, it's the stack to look at when your agent has to hold a spoken conversation in production rather than a scripted demo.
Worth watching this week: whether MCP directories start vetting what they list, now that agent skills and servers are being installed by the thousands on trust alone.
Which of these would you actually put in front of a customer today? Tell us in the comments.
Sources: last30days-skill (GitHub) · OpenMAIC (GitHub) · awesome-mcp-servers (GitHub) · Glama MCP directory · LiveKit Agents (GitHub) · LiveKit Agents docs
