Forkcast Weekly #6: agent-skills, last30days-skill, pm-skills, google-skills, superpowers

Episode Summary#

This episode covers 30 trending open-source projects from the past week. Highlights include Addy Osmani’s agent-skills framework with 59K stars, NVIDIA’s SkillSpector security scanner for AI agent skills, Google’s official GCP skills library, an 88% faster DNS tunneling VPN, and a vector database achieving 16× compression over FAISS. Also featured are a cryptographic notary for AI agent financial transactions, Apple’s native container runtime for macOS, and a 228K-star TDD workflow framework.

Repositories Covered#

agent-skills#

AI coding agents lack production-grade engineering discipline — they skip tests, produce sloppy PRs, and don’t follow industry workflows. 7 slash commands (/spec, /plan, /build, /test, /review, /code-simplify, /ship). 24 skills across 6 lifecycle phases. 229 commits. Works with Claude Code, Codex, Gemini CLI, Cursor, Copilot, Windsurf.

last30days-skill#

Google ranks pages by SEO. There’s no tool that searches actual people’s conversations across social platforms and ranks by real engagement. Cross-platform research agent skill. Sources: Reddit, X, YouTube, HN, Polymarket, TikTok, Bluesky, GitHub. Scores by upvotes, likes, and real money. v3 features: intelligent pre-research entity resolu

pm-skills#

Generic AI gives you text. Product managers need structured frameworks — PRDs, OKRs, discovery cycles, competitive analysis — not generic brainstorming. PM Skills Marketplace: 9 plugins, 68 skills, 42 slash commands. Covers full PM lifecycle: discovery→strategy→execution→launch→growth. Embeds proven frameworks (Teresa Torres, Marty Cagan, Alberto Savo

google-skills#

AI coding agents don’t know how to use Google Cloud — they need structured, authoritative skill files for GCP services like BigQuery, GKE, and Cloud Run. Google’s official agent skills for GCP. Covers BigQuery, Cloud Run, GKE, AlloyDB, Firebase, Cloud SQL, Workload Manager. Well-Architected Framework (6 pillars). Gemini API on Agent Platform. 73 commit

superpowers#

Coding agents write sloppy code without tests, skip planning, and produce unmaintainable mess — they need a mandatory TDD workflow baked in. Agentic skills framework enforcing TDD, spec-first workflows, git worktree isolation, and two-stage code review. 7-step core loop: brainstorm → plan → subagent-driven dev → test → review → finish. 228

SkillSpector#

26.1% of AI agent skills contain vulnerabilities, 5.2% show malicious intent — but there’s no security scanner for skill files before installation. NVIDIA’s security scanner for AI agent skills. 64 vulnerability patterns in 16 categories (prompt injection, data exfiltration, supply chain, MCP poisoning). Two-stage: static analysis + LLM semantic

Agent-Reach#

AI agents need to read Twitter, Reddit, YouTube, and GitHub but each platform requires separate API keys, auth flows, and fragile scrapers. Capability layer for AI agents. Zero API fees — all open-source backends. Cookie-based auth for gated platforms. Self-diagnosis via ‘agent-reach doctor’. Supports Twitter, Reddit, YouTube, GitHub, Bil

goose#

AI coding tools are siloed by provider — you’re locked into one model, one IDE, one workflow. There’s no open-source agent that works with any LLM anywhere. Open-source extensible AI agent. Desktop app + CLI + API. 15+ AI providers (Anthropic, OpenAI, Google, Ollama, OpenRouter). 70+ MCP extensions. Built in Rust (64.3%) + TypeScript. ACP protocol. Linux

Personal_AI_Infrastructure#

AI tools know nothing about your life priorities, projects, or goals — every conversation starts from zero with no persistent context about who you are. Life Operating System. Algorithm v6.3.0: 7-phase OBSERVE→THINK→PLAN→BUILD→EXECUTE→VERIFY→LEARN. 45 public skills, 171 workflows, 37 hooks. Memory v7.6 with 6 tiers. Pulse daemon on port 31337. Filesys

whichllm#

Fitting a model into VRAM is easy — knowing which of the models that fit is actually the best requires digging through stale leaderboards and guessing at quantization. Auto-detects GPU/CPU/RAM, ranks Hugging Face models by real benchmarks (LiveBench, Artificial Analysis, Aider, Chatbot Arena). Recency-aware scoring. Architecture-aware VRAM estimates. One-command cha

aisuite#

Switching between LLM providers means rewriting your integration code for every API — there’s no unified interface across OpenAI, Anthropic, Google, and others. Andrew Ng’s unified interface for multiple LLM providers. Single API across OpenAI, Anthropic, Google, and more. Simplifies provider switching and A/B testing.

tollgate#

AI agents can drain your wallet on a bad prompt — there’s no financial trust layer between agents and on-chain transactions. Cryptographic notary + Solidity guard + TypeScript SDK. Enforces YAML policies on-chain via Gnosis Safe. 3-phase architecture: Notary (Go), Guard (Solidity), SDK (TypeScript). Holds zero customer fund

container#

macOS can’t run Linux containers natively — Docker Desktop is slow, resource-heavy, and not optimized for Apple silicon. Apple’s official container runtime. Uses lightweight VMs optimized for Apple silicon. Fully OCI-compatible. macOS 26 required. Each container gets its own IP — no port-mapping. Written in Swift. 640 c

turbovec#

A 10M document corpus takes 31GB as float32 — existing vector databases are too memory-heavy for local RAG on consumer hardware. Implements Google Research’s TurboQuant algorithm. 10M docs: 31GB → 4GB (2-bit, 16× compression). Hand-written SIMD kernels (NEON/AVX-512BW) beat FAISS by 10-19% on ARM. No training phase, online inge

mattermost#

Slack and Teams lock your collaboration data in closed clouds — you can’t self-host, audit, or customize the platform for DevSecOps workflows. Open-source secure collaboration platform. Chat, workflow automation, voice calling, AI integration. Go + React. 22,673 commits, 987 releases. Self-hosted, single Linux binary + PostgreSQL. 700+ integ

maigret#

OSINT investigations require manually searching dozens of sites for the same username — there’s no automated tool to collect a full profile dossier. Username-based OSINT tool checking 3000+ sites. Recursive search finds connected usernames. Exports: JSON, CSV, PDF, HTML graph, XMind 8. Built-in web UI. Cloudflare bypass via FlareSolverr. AI analys

LMCache#

Multi-turn conversations and long-context windows burn through your LLM budget because KV caches are recomputed from scratch every request. KV cache sharing engine for LLM serving. Reduces redundant computation across multi-turn conversations. Integrates with vLLM, SGLang. C++, Python, CUDA. Apache 2.0.

MasterDnsVPN#

Standard VPN protocols are trivially blocked by deep packet inspection — DNS tunneling tools exist but are slow with large header overhead. DNS tunneling VPN. ~5-7B header overhead (88% lower than DNSTT). Up to 9x faster downloads. 6 encryption methods up to AES-256-GCM. Survived 70+ day internet blackout in Iran. 36 releases, 1,001 commi

iptv#

Finding publicly available IPTV channels across the world is fragmented — links are scattered, unverified, and constantly changing. Collection of publicly available IPTV channels worldwide. 397+ contributors, 6.4k forks. Auto-updating playlists via GitHub Actions. M3U format. EPG support via companion repo. Unlicense (CC0 equivale

music-assistant#

Streaming services are siloed — you can’t play Spotify on Sonos while browsing Tidal playlists from the same unified library. Free open-source media library manager. Connects Spotify, Tidal, Qobuz + local files. 918 releases, v2.9.0. Python 95.9%. Advanced audio analysis: BPM, key, loudness, CLAP embeddings via local ML mode

tolaria#

Your knowledge base is locked in Notion or Obsidian’s proprietary cloud. You need plain Markdown files, git versioning, and full offline ownership. Desktop app for Markdown knowledge bases. Files-first, git-first, offline-first. Tauri + React + TypeScript. 10,000+ notes workspace. 3,085 commits, 1,273 releases. MCP server for AI tooling. AGPL-3.0

openmed#

Healthcare NLP tools require sending patient data to cloud APIs — violating HIPAA. There’s no local-first, production-ready medical NER toolkit. Local-first medical NLP. 1,000+ curated biomedical NER models. 12 languages, 247 PII checkpoints. HIPAA-aware de-identification (all 18 Safe Harbor identifiers). MLX on Apple Silicon: 24-33× faster th

chatwoot#

Intercom and Zendesk cost thousands per month. There’s no open-source omnichannel support desk that handles live chat, email, WhatsApp, and social in one dashboard. Open-source omnichannel support desk. AI Agent ‘Captain’ for auto-responses. Help center portal. Custom reports and CSAT surveys. 6,285 commits, v4.14.2. Ruby + Vue.js. Self-hosted. Integrations: Slac

supervision#

Every computer vision project reinvents the same wheels — annotation, tracking, dataset splitting, format conversion. There’s no unified toolkit. Reusable CV tools: annotators, trackers, dataset loaders/splitters/mergers. Supports COCO, YOLO, Pascal VOC. Model-agnostic. 165 contributors, 4,893 commits. RF-DETR integration. MIT.

opencv#

Computer vision development requires stitching together dozens of low-level libraries — there’s no unified, battle-tested open-source CV framework. Open Source Computer Vision Library. v5.0.0 released. 1,817 contributors, 36,260 commits. C++ core with Python/Java bindings. Recent: Hough Transform 88% faster via parallel_for_. Apache 2.0.

system-prompts-and-models-of-ai-tools#

Every AI coding tool has different system prompts, internal tools, and model configurations — there’s no central reference to study how they actually work. Collection of system prompts for 30+ AI tools: Claude Code, Cursor, Devin, Windsurf, Lovable, Manus, Perplexity, Replit, v0, and more. 140k stars, 34.7k forks. 501 commits. GPL-3.0. Security research

agentsview#

You use Claude Code, Codex, and Cursor daily but have no idea how much you’re spending or which sessions are productive — agent analytics is a black box. Local-first session analytics for 25+ coding agents. 100× faster than ccusage (SQLite indexed). Token usage, cost tracking, session browser with FTS5 search. Activity heatmaps, tool usage, velocity me

career-ops#

Job hunting means spreadsheets and spray-and-pray applications — there’s no AI-powered system to evaluate, score, and track offers with integrated PDF generation. AI-powered job search system. 14 skill modes. A-F scoring (10 weighted dimensions). ATS-optimized PDF CVs + cover letters. 45+ pre-configured company portals. Batch processing with parallel sub-agents

PowerToys#

Windows lacks power-user utilities built-in — you need a dozen separate tools for window management, file renaming, color picking, and keyboard remapping. Microsoft’s official power-user utilities for Windows 10/11. FancyZones, PowerRename, Color Picker, Keyboard Manager, 20+ tools. 115k stars. MIT.

MoneyPrinterTurbo#

Creating short-form video content takes hours of editing — there’s no one-click AI tool that generates HD videos with subtitles and music. AI-powered one-click HD short video generator. Web UI + API + batch generation. Subtitle and music management. Multi-LLM support for automated content creation.

Watch#

• Video: https://www.youtube.com/watch?v=TAqm0yOFkb0
• Cover: https://i.ytimg.com/vi/TAqm0yOFkb0/hqdefault.jpg

Notes#

Transcript and notes will be added from Forkcast output artifacts.