← GitHub Repos
GitHub Repos 2026-05-15

Top 10 GitHub Repos — 2026-05-15

Top 10 GitHub Repos — 2026-05-15

Top 10 GitHub Repos — 2026-05-15

As of: 2026-05-15 Focus: AI tools, automations, agents, integrations, MCP, CLIs, and developer tooling. Snapshot: Current top 10 at the moment of generation — not a weekly recap.


Overview

This week's snapshot picks up where the 2026-05-09 digest left off and rotates entirely into projects that were either missed by the last cut or have surfaced since. The dominant theme is the substrate stack for coding agents — Claude-Code-style design tools, context-window optimisers, agent-skill cookbooks, microVM sandboxes — alongside a fast-rising autonomous-research stack (ARIS) that crossed 9,000 stars in two months. Two repos broke 1,500 stars in their first five days (nexu-io/html-anything and OpenMythos-style reconstructions still in the trending tab), and one was created the same day this digest was written and already cleared 100 stars.


1. html-anything

An agentic HTML editor bundling 75 skills × 9 surfaces (magazine, deck, poster, social-card, prototype, data report, Hyperframes) with sandboxed preview and one-click export to WeChat, X, Zhihu, HTML, and PNG. BYOK from Claude Code, Cursor, Codex, Gemini, Copilot, OpenCode, Qwen, or Aider — no first-party API key required. Created 2026-05-11 and cleared 1,900 stars in four days, the largest fresh launch in the AI-tooling tab this week.

Why it's on the list: First credible "agent-writes-the-HTML, you ship it" editor with a serious skill library and a real sandbox — fills the gap between Claude Design knock-offs and freeform vibe-coding.


2. context-mode

A context-window optimiser for coding agents that sandboxes tool output and claims a 98% reduction across 15 platforms — Claude Code, Codex, Cursor, Copilot, OpenCode, Kiro, Antigravity, Zed, and more. Distributes as a plugin / hook layer rather than a wrapper, so existing harnesses keep their workflows. Created 2026-02-23 and still adding stars at multiple-hundred-per-week pace as agent operators discover token bills are now their #1 cost.

Why it's on the list: Cost-per-task is the second axis the agent stack is now optimising, and context-mode is the most-installed open-source answer.


3. OpenMythos

A theoretical from-first-principles reconstruction of the "Claude Mythos" architecture in PyTorch / JAX, walking through attention, looped transformers, and training scaffolding using only the available research literature. Created 2026-04-18; passed 12,000 stars by 2026-05-15 on the strength of being the most readable open-source attempt at modelling Anthropic's published claims.

Why it's on the list: Strongest current pick for engineers who want to understand the post-2025 frontier-LLM stack without waiting for a paper drop.


4. Auto-claude-code-research-in-sleep (ARIS)

A markdown-only autonomous ML-research skill pack — cross-model review loops, idea generation, experiment automation, and paper writing — that runs against Claude Code, Codex, OpenClaw, or any LLM agent with no framework lock-in. Created 2026-03-10 and now at 9.4k stars and 898 forks; the "no framework, just markdown skills" approach is what is driving the velocity.

Why it's on the list: Cleanest expression of the "agents run while you sleep" pattern that's now common in academic ML workflows.


5. guizang-ppt-skill

An agent skill for generating polished HTML slide decks — editorial-magazine and Swiss layouts, image-prompt generation, social covers, and a WebGL / low-power presentation runtime — drop-in for Claude Code and Codex. Created 2026-04-23; nearly 9,000 stars and 733 forks at three weeks old as the consumer-facing "make me a deck" use case eats LangChain-flavoured pipelines.

Why it's on the list: Best-in-class example of a single skill package outshipping a framework for a narrow, high-value workflow.


6. open-codesign

An open-source Claude Design alternative — Electron desktop app, prompt-to-prototype / slides / PDF flow, multi-model fanout (Claude, GPT, Gemini, Kimi, GLM, Ollama), BYOK, local-first, MIT. Imports your existing Claude Code or Codex API key with a click. Created 2026-04-18 and a near-mover of the "Claude Design clone" wave.

Why it's on the list: Cleanest, most-installable answer to "I want Claude Design but I want to bring my own model and run it locally."


7. Claude-Code-Design-AI

A Claude Code plugin pitching itself as the AI UI/UX architect — screenshot-to-React, Figma-component generation, Tailwind / shadcn-ui scaffolding, wireframe rendering, SVG icon creation, dark-mode toggles, and responsive-layout tooling. Created 2026-05-13; cleared 370 stars in two days as the "Claude Design alternative" arms race kept compounding.

Why it's on the list: Frames itself as the front-end-export bridge between Claude Design and a real React codebase — a missing piece in most of the clones.


8. harmonist

A portable AI-agent orchestration framework with mechanical protocol enforcement — 186 agents, zero runtime dependencies, Python-only, designed to be dropped into Claude Code or Cursor without a sidecar. Created 2026-04-23 and crossed 1,600 stars / 342 forks by 2026-05-15, with the protocol-enforcement angle differentiating it from the LangGraph / CrewAI cohort.

Why it's on the list: Strongest "single-binary multi-agent" pick this week — useful when you want an agent council without spinning up an event bus.


9. journal-adapt-writing-skill

A Claude-Code skill that learns a target journal's writing conventions from its published papers, then rewrites a manuscript section-by-section to match. LaTeX-aware, economics-flavoured by default, MIT-licensed. Created 2026-05-13 and at 233 stars by 2026-05-15 — small in absolute terms, but a sharp instance of the "domain-specific skill" pattern eating prompt-engineering posts.

Why it's on the list: Concrete proof that narrow, single-task skills outperform generic "academic writing" prompts when the workflow is well-defined.


10. forkd

A Rust-based microVM sandbox built on KVM that forks a warmed parent in 101 ms using copy-on-write snapshots — aimed squarely at agents that need a clean OS for every tool call. Created 2026-05-11. Sits in the same problem space as Firecracker / Mirage but trades ergonomics for raw fork latency.

Why it's on the list: Lowest-latency open-source sandbox primitive for agents that have to spin up a clean kernel per task — paired with mirage (last week's #1) it forms a credible agent-runtime substrate.


Honorable Mentions (Optional)

  • DenisSergeevitch/agents-best-practices — Provider-neutral agent-skill best-practices repo, 115 ★ on its release day (2026-05-15).
  • Evokoa/pgGraph — Open-source graph-DB extension for existing Postgres, pitched at agent-relationship workloads (137 ★, created 2026-05-13).
  • sno-ai/llmix — Production LLM call layer with cache, retries, circuit breakers, key rotation, and Python/TypeScript/Rust parity (128 ★, created 2026-05-09).
  • Storybloq/storybloq — Cross-session context for Claude Code via CLI + MCP server + /story skill (510 ★).
  • oh-story-claudecode — Chinese web-novel writing skill pack with end-to-end story / cover / de-AI-ification flow (1,123 ★).

Sources

More from GitHub Repos