A strictly vetted directory of open-source artificial intelligence repositories, local LLM engines, multi-agent frameworks, and vector search tooling. Annotated with personal engineering takes from production builds.
Curated Open-Source AI Repositories
garrytan/gstack ↗
23 Markdown slash commands plus 8 power tools staging a full sprint — CEO, eng manager, QA, security officer — from idea to shipped PR with a 30-second install for Claude Code. When NOT to use: skip if your team already runs a release pipe — overlapping personas double-process small teams; adopt one role at a time.
“I run a small team from Junagadh, so a staged sprint in slash commands fits how we already ship. The QA-lead command caught a broken UPI retry path before a Surat launch — P95 held 780ms because the checklist forced a load pass. I trimmed the CEO persona on day one; strategy prompts without profit-and-loss context generated noise for my fixed-price quotes.”
samber/cc-skills-golang ↗
Go-specific agentic skills for coding assistants: conventions, testing, security, observability, architecture, framework guides. Human-edited and tested with a claimed plus-41-point eval gain. When NOT to use: skip for single-language Python shops — loading Go conventions into context burns tokens with zero return outside Go repos.
“I keep one Go service (the telemetry collector fronting my ledger) beside a PHP and Python stack, and generic agents write unidiomatic Go without guardrails. The testing skill standardized table tests across that service in one sitting from Junagadh. I load these only inside the Go repo via directory-scoped config so Python work never pays the context tax.”
1jehuang/jcode ↗
JCode — the most RAM-efficient harness. Rust, 12.8k stars, +2.5k week Sep 18 2026.
“I test minimal harnesses on a Rs 6K VPS where every extra dependency breaks deploys, and RAM efficiency is the metric nobody benchmarks. JCode claiming the leanest harness in Rust is the exact shape I would run on Pi 5 edge agents at 62 tok/s beside GenericAgent. I would measure its context use against my 30K-context discipline before trusting the claim.”
ConardLi/garden-skills ↗
Garden Skills — open-source skills collection: web design, knowledge retrieval, image generation and more. 12.5k stars (verified via API Sep 19 2026).
“I maintain a 41-skill Junagadh catalog where each entry earns its place through client revenue. Garden-skills spanning design plus retrieval plus image generation is the cross-domain shelf I would audit skill by skill — keep the retrieval entries that beat my pgvector 42ms baseline, drop the rest. I install with tight approvals after the broad-permission audit, same gate as every skill.”
cursor/plugins ↗
Cursor plugins — official plugin specification plus plugins. TS, 6.3k stars, +1.4k week Sep 18 2026.
“I wire Cursor seats for Gujarat SME pilots where repeatable deploys beat hand-rolled prompts. The official plugin spec is what I would pin every Cursor setup to — read the spec, ship one verified plugin per seat, refuse the rest. I test new entries against my Rs 6K VPS harness before they touch client repos.”
humanlayer/skills ↗
HumanLayer skills — approval workflows for AI agents. TS, 3.1k stars, trending Sep 2026.
“I gate every money action behind a two-tap owner queue in Junagadh — the deny-by-default pattern that closes deals. HumanLayer productizing that exact approval layer as skills is the upstream I would track closest; if their contact-grant flow beats my n8n wait-node, I adopt it. My staging once fired duplicate reversals without a gate — never again, any vendor.”