Open Source Radar — September 17: voice, audits, and MCP testing
Today's trending board is mostly agents and the skill packs that feed them, so the interesting movement is one layer down — the tools that give an agent a voice, test the servers it talks to, and check whether the code it just wrote is actually safe.
Voicebox (TypeScript, ~54,700 stars, MIT) — A local-first voice studio that covers both halves of the voice loop: cloning and speech generation on the way out, dictation into any text field on the way in, which is why it reads as an open answer to ElevenLabs and WisprFlow at once. Seven speech engines ship in one app — Qwen3-TTS, Chatterbox in multilingual and fast variants, Kokoro, HumeAI's TADA and others — across 23 languages and 50-plus preset voices, with per-profile personas and an audio-effects chain. The feature worth the install is the MCP server: any MCP-aware agent can speak to you in a voice you cloned, from models and voice data that never leave your machine. Desktop builds cover macOS and Windows with a Docker path for Linux; it's Tauri rather than Electron, and prebuilt Linux binaries aren't there yet.
Cloudflare security-audit-skill (JavaScript, ~8,700 stars, MIT) — The skill that seeded Cloudflare's internal vulnerability-discovery harness, published as a single-repo starting point. It runs six phases: map the code's architecture and trust boundaries into a coverage ledger, assign isolated hunters to ledger units, hand every candidate to a fresh verifier whose only job is to disprove it, write the survivors into a machine-readable findings file, then have separate agents re-verify the final source claims before anything is reported. Three verdicts stay strictly apart — confirmed, needs validation with an exact unresolved fact and no severity, and rejected — and the programs that validate the file are the boring deterministic kind rather than another model. Two design rules do the real work: defense-in-depth gaps are not vulnerabilities, and severity requires demonstrated impact. The honest limit is that without an OS-enforced sandbox blocking outbound network access, it won't execute target code at all and files the leads as unresolved. Cloudflare also reports that one run found roughly half the vulnerabilities that repeated runs found in total — the best argument yet for auditing on a schedule instead of once.
MCPJam Inspector (TypeScript, ~2,200 stars, Apache-2.0) — An open testing and evaluation platform for people who write MCP servers, aimed at the debugging problem the MCP boom created: ChatGPT, Claude, Cursor and Copilot each read the same server differently, and nobody notices until a user does. It exercises your tools, prompts, resources and authorization against 16 client configurations and more than 170 models, with a trace of every JSON-RPC message, a guided OAuth conformance checker covering four protocol versions including the current draft, and a widget emulator that switches between desktop, tablet and mobile so an app's UI can be tested without a host client. Evals score tool-call accuracy over time and can gate a pull request in CI, and teams share server groups in one synced workspace so everyone tests the same config. Apache-2.0, on Product Hunt today, and actively moving — thousands of commits with the newest landing this morning.
Knowledge Work Plugins (Python, ~24,400 stars) — Anthropic open-sourced 11 plugins that turn Claude into a specialist for a job function: sales, marketing, legal, finance, data, product management, customer support, enterprise search, bio-research and more, each bundling the skills, connectors and explicit commands for that role. Every component is markdown and JSON — no code, no build step — and the connectors are the substance, wiring the assistant to the CRM, data warehouse, ticket queue or research database the role actually lives in. The framing inside the repo is unusually plain: these are generic starting points, useful mainly as the base you edit, swapping connectors and dropping in your company's terminology and processes. It's trending because it settles where agent configuration gets packaged now, and because the same repo includes a plugin for building your own. Worth a look if your organization's "AI setup" is still somebody's personal chat history.
Worth watching this week: whether MCP testing and agent-run security audits become routine CI steps — or stay the thing teams do by hand after something breaks.
If your agent audited its own codebase and handed you a machine-readable verdict, would you trust it or insist on a second human pass? Tell us in the comments.
Sources: Voicebox (GitHub) · Cloudflare security-audit-skill (GitHub) · MCPJam Inspector (GitHub) · Knowledge Work Plugins (GitHub)
