Curated Open-Source AI Repositories

“I run Curro AI content studio + Warp agent env for production — claude-studio-toolkit brings that studio UX to Claude Studio for MLOps. Its skill + workflow pack is the studio bridge I test alongside my day-in-life 06:00-22:00 routine.”

“I maintain 41 skills where docs rot kills agent accuracy — openwiki auto-generates AI-consumable docs so my MCP tools stay discoverable. Its agent-first docs format is what I enforce before any skill enters my Junagadh catalog.”

“I run client calls where DPDP forbids cloud recording — meetily transcribes 4x faster locally with Ollama summaries, staying inside VPC. Its Rust + Parakeet stack is the in-VPC meeting layer I deploy for Rajkot/Surat SMEs.”

“I automate Gujarat SME paperwork (GST sheets, RFQ docs) where agents choke on binary Office formats — OfficeCLI gives agents a single-binary Word/Excel/PPT interface with no Office license. Its agent-native document API is the missing tool in my n8n + MCP fan-out.”

“I run Claude Code + Codex + OpenCode harnesses where cross-harness delegation is manual — codex-plugin-cc lets Claude Code delegate review tasks to Codex natively. Its harness-bridge pattern is the interop layer my Junagadh multi-harness setup needed.”

“I test Google ADK vs LangGraph vs CrewAI for Gujarat subagents — munder-difflin is the local-first multi-agent harness that runs the same supervisor pattern without cloud lock-in. Its TS harness is the lightweight alternative I benchmark against ADK from Junagadh.”

“I gate every tool via OPA + HITL at the runtime layer — substrate builds that substrate as a Go core system instead of Python glue. Its systems-level agent runtime is the foundation pattern I study for Pi 5 edge agents at 62 tok/s.”

“I maintain a 41-skill catalog where 26% of trending skills request broad permissions per my SkillSpector audit — agent-skills validates every skill before install, the same sign-before-ship gate in my Junagadh catalog. Its cross-harness registry (Claude Code to Copilot) is the distribution shape my library pages follow.”

“I ship WhatsApp-first automation for Gujarat SMEs where 98% opens beat 12% email — DeskcommCRM bakes AI agents plus WAHA WhatsApp directly into a self-hosted CRM, the exact stack I assemble with n8n + MCP for Rajkot RFQ inboxes. Its sell-by-chat OS is the fastest path from lead to UPI link I have seen open-sourced.”

“I benchmark Phi-4 Mini 3.8B at 300 tok/s on 3GB where every megabyte decides deployability — needle compresses a foundation model to 14MB for wearables and phones, an order beyond my Pi 5 edge work. Its tiny-device target is the endgame of my on-device thesis for 4G-drop Junagadh field agents.”

“I ship browser-use automation where agents stall on Cloudflare and logins — Cloudflare giving agents a first-class computer primitive removes the whole stealth-browser arms race. Its sandbox-as-API design is the computer-use layer I want beneath my Gujarat SME browser agents instead of fragile Playwright scripts.”

“I installed this in my Junagadh lab after agents kept dumping 400-line answers on small fixes. Responses got scannable in one day. It pairs well with my P95 ledger reviews where I skim 20 agent runs fast.”

“I test team CLIs for Gujarat SME rollouts where 5-10 staff share one agent setup. This one standardizes the entry point instead of everyone hand-rolling prompts. I run it on a Rs 6K VPS with shared Valkey cache.”

“I keep an eye on CAD skills for Rajkot engineering clients who quote sheet-metal jobs. Drafting a bracket concept in minutes instead of hours changes quoting speed. I verify outputs against FreeCAD before showing clients.”

“I ran the paper-trading loop for a Surat client who wanted research summaries, not auto-orders. The debate logs read like an analyst desk. I cap it at end-of-day reports on the VPS, never live execution.”

“I point Junagadh interns here before touching my client stack. Rebuilding tokenizers and retrievers by hand removes magic. Two interns shipped MCP tools within a month after this.”