kepano/obsidian-skills (GitHub) - 45,636 stars for teaching your agent to speak Obsidian's language: Markdown, Bases, JSON Canvas, all through the CLI. Most agent integrations with knowledge tools are bolted on and brittle. This one respects the open formats Obsidian already uses. That means it survives an update instead of breaking on one. If you keep notes in Obsidian and want an agent that can actually read and write them properly, install this first.
cathrynlavery/diagram-design (GitHub) - Twenty nine diagram types built specifically for Claude Code, self-contained HTML and SVG, no Mermaid in sight. Over 14,000 stars and 4,500 of those landed in a single day, which tells you how badly people wanted this. Every developer who's watched Claude generate a diagram that looks like it was drawn by a committee will get the appeal instantly. Clean output, no shadows, no dependency hell. Install it.
superset-sh/superset (GitHub) - Run over 100 coding agents in parallel, Claude Code or Codex or any CLI agent you fancy, each one sandboxed in its own isolated environment. 12,889 stars and climbing, written in TypeScript. This solves a real problem: agents stepping on each other's file changes when you try to run more than one at once. Worth knowing about even if you never need a hundred agents running. Ten will do most of the damage you actually want.
calesthio/OpenMontage (GitHub) - What happens when you point agentic AI at a genuinely hard domain instead of another chatbot wrapper. OpenMontage claims to be the first open-source agentic video production system, and it backs that claim with twelve production pipelines and over 100 tools, all in Python. Video is a brutal place to build agent workflows. Failure is visible and expensive to fix. If it holds up under real use, it's the most ambitious thing on this list by some distance.
triggerdotdev/trigger.dev (GitHub) - A TypeScript framework for building and deploying fully-managed AI agents and workflows, sitting at 15,825 stars. Less flashy than most things on this list. Trigger.dev is infrastructure, not a demo. If you're past the "look what Claude can do" stage and into "this needs to run reliably every night without me watching it", this is the layer that gets you there.
Look at what actually got starred this week. Not another agent framework promising to orchestrate your life. Tools that plug an agent into something that already existed and worked: Obsidian's own formats, Claude Code's own output, a CLI you already run. The agent doesn't run the show any more. It's the connective tissue.
That matches what I keep telling clients when they ask why their AI project stalled after the demo. The demo works because a demo is scoped. Production fails because production means fitting into fifty things that already exist in the business, none of which were built with an agent in mind. OpenMontage is the outlier here, a system built from scratch for a hard domain, and it's the riskiest bet on this list precisely because of that. The safer money this week was on integration, not invention.
Which of the five would you actually run in production next week?
Tell me in the comments.
Forward this to one person who should be using AI better than they are. Reply with what you built, tried or broke this week. I read every one.
Gareth, founder of The Anthropic Stack (theanthropicstack.com)
P.S. Shipping something this week? Railway gives Anthropic Stack readers $20 in credits to start: theanthropicstack.com/railway. It is a referral link, so I earn a small cut if you stay. I only point you at tools I run myself.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.