Open Source Radar — October 1: the tools agents run on
Today's trending list is less about new models and more about the plumbing: what agents can see in your codebase, which skills they can load, and what they can produce once you point them at a blank canvas.
MoneyPrinterTurbo (Python, ~127,800 stars) — A one-stop AI short-video factory: give it a topic or a keyword and it writes the script, picks stock footage, generates subtitles and background music, then cuts a finished HD clip. It offers four entry points — an agent interface, a web UI, an API and a CLI — and supports a long roster of script models from Kimi and OpenAI to DeepSeek and local Ollama, plus a dozen text-to-speech providers including a free option that needs no API key. Batch mode takes a manifest of up to 100 tasks and reports results in one summary, which is the difference between a demo and a content pipeline. It is the single hottest AI repo on trending today, and the honest use case is high-volume short-form output where per-video quality matters less than throughput.
CodeGraph (Rust kernel, ~72,700 stars) — A pre-indexed knowledge graph of your codebase that keeps itself fresh, built for how coding agents actually work: it parses 20 languages with tree-sitter, resolves cross-file references into a local SQLite graph, and watches the project with OS file events so the index never goes stale while an agent edits. The payoff is fewer tokens and fewer tool calls — one explore call returns the relevant source, the call flow and the blast radius instead of an agent spawning a search sub-agent and reading files one by one. It wires into Claude Code, Cursor, Codex, Gemini CLI, Copilot and more through MCP, runs 100% local, and is MIT-licensed. If your agent keeps re-reading the same files to answer one question, this is the class of fix that pays for itself.
modelcontextprotocol/servers (TypeScript, ~90,900 stars) — The MCP steering group's own collection of reference servers — filesystem, git, memory, fetch, sequential thinking, time — and the official showcase of what the protocol's SDKs can do. It is on trending again as a reminder of where the ecosystem actually lives: the repo now explicitly points readers to the MCP Registry for the full server catalog and keeps only the educational examples, with a plain warning that these are demos rather than production-ready software. Reach for it when you are building your own server and want the canonical pattern, not as a dependency you ship. Ten SDKs in ten languages are listed here, which tells you how standardized the wiring has become.
awesome-claude-skills (curated list, ~76,300 stars) — A catalog of more than 1,000 production-ready Claude Skills spanning document processing, coding, data work, security and communication — usable from Claude.ai and Claude Code, and by Codex, Cursor, Gemini CLI and other agents that adopted the open skill format. The repo doubles as the best short explainer of how skills differ from MCP and tools: MCP grants access, tools perform actions, and skills define the workflow — and because skills load progressively at roughly 100 tokens of description each, one agent can carry hundreds without bloating its context window. Treat it as a directory rather than a dependency; individual skills carry their own licenses. Useful whenever you are deciding what to teach an agent next.
Worth watching this week: whether the code-graph layer becomes standard equipment for coding agents, the way MCP wiring did a year ago.
Which of these is actually earning its place in your workflow — and which is trending noise? Tell us in the comments.
Sources: MoneyPrinterTurbo (GitHub) · CodeGraph (GitHub) · MCP servers (GitHub) · awesome-claude-skills (GitHub)
