Open Source Radar — October 26: Claude SEO, Microduck, VoiceStudio
Agent skills and coding harnesses own GitHub's trending page this morning, and for once the most interesting entries are replacing a bill you already pay.
claude-seo (Python, 21,400 stars, MIT) — A free SEO audit that runs inside Claude Code, built from 26 sub-skills and 19 specialist agents that fan out across a site and come back with a prioritised fix list in 10 to 15 minutes, per its README. It covers the boring half — crawlability, indexing, sitemaps, redirects, Core Web Vitals — and the half most audit tools still skip: whether Google's AI Overviews, AI Mode, ChatGPT, Perplexity and Gemini are likely to cite your pages at all, plus agent-readiness checks for llms.txt and WebMCP. The pitch is aimed at freelancers and agencies who want an audit on the first sales call, and the author's own site reports 10.8K Search Console clicks in its first 12 weeks. Treat that number as marketing, then point it at a site whose traffic you already know.
microduck (C, Rust and Python, ~12,000 stars, Apache 2.0) — Pollen Robotics open-sourced the brain of a 25 cm, 800 g two-legged duck that walks on reinforcement-learning policies: a 50 Hz control loop driving fifteen servos on a Rockchip RK3566, with the camera and radios on top. The split is the interesting part. This repo is the runtime; the policies are trained next door in MuJoCo and PPO and exported to ONNX, so a new walking brain drops in without touching the robot. Put wheels on and it loads a different policy and rolls; knock it over and it stands back up. Updates are verified, health-gated and reversible, and a simulator runs the real control daemons against a simulated body for anyone without a duck on the desk.
VoiceStudio (Python, ~87,200 stars, AGPL-3.0) — The open-source ElevenLabs replacement now runs on a phone. Cloning, voice design, dubbing and transcription across a claimed 646 languages, on your own hardware — and the beta published on October 23 adds preview iOS and Android builds that keep working offline once the models are downloaded, per its release notes. It still exposes a local API and MCP server so an agent can drive it directly. Two caveats the project's own docs are honest about: 646 is a catalogue across engines rather than quality in every language, and each speech model carries its own licence even though the app code is AGPL.
thinkrail (TypeScript, ~2,100 stars, Apache 2.0) — JetBrains' incubator answer to the agentic IDE, built around the open pi coding agent rather than a model of its own. Every workspace is a git worktree with its own branch, so several agents can work in parallel without touching your checkout, and the top bar tracks each branch against origin and its pull request with CI status one click away. The bet is durability: project facts and reusable skills live in living specs, and full extension support is next so the IDE can grow tools from how you actually work. At 2,100 stars and 215 open issues it is early — the 0.3.0 nightlies shipped this week — but it is the first JetBrains project that treats the coding agent as the platform instead of a plugin.
Worth watching this week: whether these agent skills keep their value once the harnesses ship the same features themselves.
Which of these would you actually install, and which is a wrapper around a prompt you could write yourself? Write to us via our contact page.
