The AI repositories that gained the most GitHub stars this week, each introduced well enough to tell whether it's worth trying: what it does, what it takes to run, and how mature it is. Written from each README and the GitHub API only.
A collection of agent skills the author says he uses daily. It includes grilling sessions where the agent questions you about requirements (/grill-me, /grill-with-docs), spec and ticket writing (/to-spec, /to-tickets), implementation (/implement), test-driven development (/tdd), bug diagnosis (/diagnosing-bugs) and architecture review (/improve-codebase-architecture). Unlike process-owning approaches such as GSD, BMAD and Spec-Kit, the skills are meant to be small and composable, according to the README.
Skills split into user-invoked ones you type yourself and model-invoked ones the agent can reach for when a task fits.
Before you try it
Installing both the plugin and via skills.sh gives you every skill twice, so pick one per agent.
Run /setup-matt-pocock-skills once per repo before using the engineering skills; it sets the issue tracker, triage labels and docs location.
The README says /improve-codebase-architecture is a survey that surfaces candidates; it won't untangle an old codebase for you.
Voice & video · TypeScript · this week +4,081 · ★ 60.9k
HyperFrames is an open-source framework that turns HTML, CSS, media, and seekable animations into deterministic MP4 videos. You can use it locally through the CLI, from AI coding agents via skills, or as the rendering core of hosted authoring workflows.
It ships a Claude Code plugin and skills installable with `npx skills add`. The README says these work with Claude Code, Codex, Cursor, Gemini CLI, IBM Bob, and other agents that support skills. Animation runtimes covered include GSAP, Lottie, Three.js, Anime.js, CSS, WAAPI, and TypeGPU, and the CLI has commands for HeyGen-hosted cloud rendering and AWS Lambda rendering.
Before you try it
There are 21 skills, and a non-interactive `skills add` run without `--skill` installs all of them. The README points to `npx hyperframes skills update` for the core set only.
`skills add` resolves the skills.sh registry, which the README says can lag `main` by hours. For the newest copy, it recommends `npx hyperframes skills update`.
A persistent memory system for Claude Code. It captures tool-usage observations automatically, generates semantic summaries, and supplies them to later sessions. It uses lifecycle hooks, a local Bun-managed HTTP worker, a SQLite database and a Chroma vector database for hybrid semantic and keyword search.
It installs through the Claude Code plugin marketplace, and the npx installer also targets OpenCode, T3 Code, Antigravity CLI, OMP, Pi, DeepSeek Harness and Grok Bot. A separate install script covers OpenClaw gateways. Search runs through three MCP tools (search, timeline, get_observations) in a layered workflow that the README says saves about 10x tokens by filtering before fetching details.
Before you try it
The default installer asks you to sign in to claude-mem in a browser (email magic link). Signing in starts a 30-day free trial of the claude-mem observer, after which memory falls back to your Anthropic plan unless you subscribe. Passing --provider, setting CLAUDE_MEM_ONLINE_OPTIN=false, or running in CI or a non-interactive shell skips sign-in.
`npm install -g claude-mem` installs only the SDK/library and does not register plugin hooks or the worker service. Use npx claude-mem install or the /plugin commands instead.
The Grok Bot integration has no host hooks, so it watches chat log files. It defaults to hosted memory (CMEM Pro), and the local observer is opt-in with --provider host.
This plugin gives an agent local workflows for generating 3D models as STEP, GLB, STL or 3MF files. The README says it also runs design-for-manufacturing checks, generates engineering drawings, and connects to 3D printing, sheet metal and CNC fabrication services.
It works with agents that support plugins or the skills framework, including Claude Code, Codex, Cursor, Gemini and Grok. Claude Desktop connects through an MCP server (cadgen mcp), and agents without a plugin can install just the skills and view models in a browser-based CAD Viewer. Under the hood it runs the cadgen package, built on build123d and Open CASCADE, through uv.
Before you try it
Usage stats and crash reports are sent by default, tagged with a random ID. The README says file names, paths, contents and prompts are never included. Turn it off with `uvx cadgen telemetry off`.
On Windows 11 with Smart App Control enabled, the unsigned native module in OCP is blocked, so cadgen commands fail with `DLL load failed while importing OCP`.
It makes one anonymous version-check request at most once a day; `CADGEN_UPDATE_CHECK=0` disables it.
OpenRig is open-source software that runs AI coding agents as a persistent, organized team instead of a pile of terminal sessions. You define the team in YAML and boot it with one command; you talk to a lead agent about the outcome you want, and it coordinates specialists.
Claude Code and Codex can run in the same rig, managed as one system. It runs on tmux and can use Herdr or Ghostty for terminal views. It ships three starting teams: starter (a Claude Code builder and a Codex reviewer), workshop, and factory (seven agents).
Before you try it
Setup and operation write to files such as ~/.tmux.conf, ~/.claude/skills, ~/.agents/skills, and Codex hook configuration and trust records. The README advises backing up the relevant files before installing.
Native Windows is not supported and WSL2 has not been tested. On Apple silicon Macs, Node.js 22 is required.
You need an existing signed-in account for at least one of Claude Code or Codex; the README says no second subscription is required.
This is Cursor's plugin marketplace repository. Each plugin is a standalone directory at the repo root with its own `.cursor-plugin/plugin.json` manifest, and a root `.cursor-plugin/marketplace.json` lists them all. Plugins contain agent skills as SKILL.md files with frontmatter.
Developer-tool plugins include branch review (thermos), PR Review Canvas, plugin scaffolding (Create Plugin), fan-out across parallel cloud agents (Orchestrate), a Cursor SDK plugin for building with the TypeScript SDK, skills for the Cursor Origin API, and Grok voice for apps. The third_party folder holds integrations for services such as Gmail, Google Drive, GitHub, Playwright, Salesforce, HubSpot, BigQuery and Excalidraw, plus several brokerage and finance services.
Before you try it
Most third-party integrations appear to need an account with the external service (Gmail, Salesforce, Gong, etc.), though the README doesn't describe the auth method.
Some plugins can prepare or execute trades: Robinhood works in an Agentic account, and Coinbase can preview or place trades, per their descriptions. eToro and Interactive Brokers are also listed.
Some plugins list individual authors such as Eric Zakariasson, Lauren Tan and Dylan Gattey. The README doesn't state a license.
This Claude Code skill rewrites dense, ambiguous English following ASD-STE100, the controlled-language standard from aerospace and defense. It targets text an AI agent must parse with no human to ask: agent output, tool descriptions, error messages, and inter-agent instructions. It flags violations sentence by sentence (passive voice, present perfect, long sentences, phrasal verbs, nominalizations, semicolons, marketing adjectives) and rewrites them without dropping facts, conditions, or scope qualifiers.
It picks Strict mode for procedures, error messages, and tool descriptions, or STE-flavored mode for READMEs and PR descriptions, which relaxes the fixed-vocabulary rules. By default it returns only the rewritten text; asking for "show the diff" gives a before/after table naming the rules.
Before you try it
It does not include ASD's official ~900-word approved dictionary. It applies the standard's principles only, so certified STE-compliant documentation requires the real standard.
The bundled linter checks structural patterns only. It does not verify that a rewrite preserves meaning, does not check noun-cluster length, and passes phrasal verbs outside its short list (e.g. "take off the panel").
The npx skills CLI sends anonymous install telemetry; set DISABLE_TELEMETRY=1 to opt out. The README says it is not built for creative writing or marketing copy.
It sits as a proxy between an agent and its real model API, rewriting Anthropic/OpenAI streams with acp-kernel compression. The model decides when and what to compress; the proxy injects four tools (compress, decompress, search_context, acp_status) that the model calls itself. Summaries are written in small ranges, can be restored on demand, and keep the cache prefix intact, per the README.
It supports many clients, including Claude Code, Codex, OpenCode, pi, Gemini CLI, Kimi, Qwen Code, GitHub Copilot CLI, aider, goose and Zed. Depending on the client, integration is a `bili ` launcher, a `bili plugin install ` native plugin, cert-MITM, or a `/bili/` URL prefix. The README says a healthy session keeps a 95–97% prefix-cache hit rate, with compression costing 2% or less.
Before you try it
Integration modes differ per client. jcode, mcode, aider, copilot, amp, crush and zed have no native plugin mode and use a launcher or cert-MITM only, per the README.
The codex native plugin (MCP-shell tools) requires starting bili first; the shell never spawns a proxy and never routes codex's own traffic.