The Pull: Week 22, 2026
This week in open source: Maps for messy code, translators for stubborn docs, and some much-needed rules for AI design
1) Understand Anything: Codebases Need Maps, Not Swagger
github.com/Lum1104/Understand-Anything | License: MIT
Featured in The Push: May 24th, 2026
Understand Anything turns codebases and docs into an explorable knowledge graph with search, guided tours, and change impact context. What makes it clever is the hybrid architecture, deterministic structure plus LLM interpretation, which makes the graph useful instead of decorative. Teams doing onboarding, product discovery, diligence, or AI-assisted coding should look.
3205 stars/day | 8.3 watchers/day
2) MarkItDown: Your Documents Need a Translation Layer
github.com/microsoft/markitdown | License: MIT
Featured in The Push: May 30th, 2026
MarkItDown turns PDFs, Office docs, media, and web content into Markdown that models can actually reason over. The smart part is the routing architecture plus Markdown as the common output layer, which makes ingestion consistent across wildly different file types. Anyone building AI around documents should look.
3044 stars/day | 8.0 watchers/day
3) Taste Skill: AI Design Finally Grows Standards
github.com/Leonxlnx/taste-skill | License: MIT
Featured in The Push: May 25th, 2026
Taste Skill turns design taste into an installable layer for coding agents, with specialized skills for frontend generation, redesigns, and image-first workflows. The clever part is the modular architecture, which treats visual judgment like reusable infrastructure instead of prompt superstition. Founders, agencies, and AI product teams should look.
1733 stars/day | 2.5 watchers/day
4) ECC: The Ops Layer Agents Needed
github.com/affaan-m/ECC | License: MIT
Featured in The Push: May 25th, 2026
ECC turns AI coding from a bunch of disconnected prompt rituals into a portable operating layer with reusable skills, memory, security, and cross-harness consistency. The clever bit is packaging behavior so it survives tool switching. Teams running multiple AI workflows, or building products on top of them, should look closely.
1366 stars/day | 7.0 watchers/day
5) Knowledge Work Plugins: AI Needs Job Training
github.com/anthropics/knowledge-work-plugins | License: Apache-2.0
Featured in The Push: May 24th, 2026
Knowledge Work Plugins turns Claude into a role-specific coworker by packaging workflows, commands, and system access into editable plugin bundles. The clever part is the packaging, company process becomes portable software. Teams standardizing support, sales, product, finance, or research should look closely.
685 stars/day | 2.3 watchers/day
6) Harness: AI Teams, Not Prompt Piles
github.com/revfactory/harness | License: Apache-2.0
Featured in The Push: May 30th, 2026
Harness turns a plain-language domain description into an AI team with roles, coordination patterns, and reusable skills. The clever part is that it treats team architecture, not prompting, as the product. Worth a look for operators, agencies, and anyone building repeatable AI workflows.
373 stars/day | 4.0 watchers/day
7) Pi: The Agent Stack, Unbundled
github.com/earendil-works/pi | License: MIT
Featured in The Push: May 24th, 2026
Pi turns the messy parts of agent products, model routing, session state, tool execution, and interface rendering, into one coherent stack. The clever bit is the persistent runtime underneath multiple surfaces. Worth a look for anyone building AI products that need to feel consistent, controllable, and durable.
625 stars/day | 1.9 watchers/day
8) Claude Code: The Terminal Wins Again
github.com/anthropics/claude-code
Featured in The Push: May 30th, 2026
Claude Code turns the terminal into an AI work surface that can read code, take actions, and follow local rules. What stands out is the hook and plugin architecture, which makes the assistant governable, not just powerful. Best for teams that want AI inside real software workflows, with guardrails.
547 stars/day | 4.0 watchers/day
9) Free Domain: ICANN’s Price Tag Looks Optional
github.com/DigitalPlatDev/FreeDomain | License: AGPL-3.0
Featured in The Push: May 26th, 2026
Free Domain turns internet naming into a free utility by letting people claim managed domains and connect them to outside DNS providers. The clever bit is the unbundling: namespace access stays simple while infrastructure stays portable. Founders, creators, student groups, and anyone testing ideas online should look.
1187 stars/day | 5.8 watchers/day
10) Stop Slop: AI Writing Needed an Editor
github.com/hardikpandya/stop-slop | License: MIT
Featured in The Push: May 26th, 2026
Stop Slop turns AI writing cleanup into a reusable editing system. The clever part is the layered design: phrase bans, structural constraints, and a scoring rubric that gives models a real feedback loop. Anyone publishing AI-assisted prose, especially founders, marketers, and students, should look.
565 stars/day | 1.8 watchers/day
11) Jellyfin: Plex’s Rent-Free Counterattack
github.com/jellyfin/jellyfin | License: GPL-2.0
Featured in The Push: May 26th, 2026
Jellyfin turns personal media libraries into a polished, self-hosted streaming service. What makes it interesting is the server-first architecture, especially the way transcoding, metadata, and device compatibility are handled as core product logic. Worth a look for anyone with a NAS, a messy archive, or a distrust of rented convenience.
88 stars/day | 0.4 watchers/day
12) Cmux: The Terminal Finally Understands Interruptions
github.com/manaflow-ai/cmux | License: Other
Featured in The Push: May 25th, 2026
Cmux turns a macOS terminal into a control surface for parallel AI coding sessions, with pane-level notifications, metadata-rich sidebars, and a built-in browser that agents can actually use. The clever part is attention routing, not terminal theming. Worth a look for anyone juggling multiple agents, remote sessions, or fast-moving software projects.
195 stars/day | 0.0 watchers/day