Seongwoo Choi / Journal

Guide · AI agent tools

AI agent repos on GitHub: 23 tools to compare

This guide compares 23 selected repositories for coding agents and related AI applications. They include instructions, harnesses and tools for Claude Code, Codex and OpenCode. The selection is editorial; it is not a comprehensive ranking of GitHub's fastest-growing projects.

Star counts are a dated snapshot from the GitHub API. They measure attention, not quality or recent growth. Read the purpose and check notes before choosing a project, then verify the current README and release history.

Ranking by stars

This selection is ranked by total stars, not measured recent growth. Average stars per day is total stars divided by repository age; it is not a daily time series. The 7-day date delta appears only with snapshots dated exactly seven days apart; capture times within those days may differ.

#RepoGroupStarsAvg/day
1REANew capabilities437,2242,122
2Matt Pocock SkillsMethodology and skill packs316,4411,147
3SuperpowersMethodology and skill packs313,481798
4ECCMethodology and skill packs298,4111,022
5DeepSeek HarnessHarnesses and control surfaces274,7853,233
6Karpathy-inspired guidelinesMethodology and skill packs230,194813
7OpenCodeHarnesses and control surfaces225,000405
8PonytailMethodology and skill packs185,7271,263
9GraphifyMemory and codebase context143,409661
10OrcaHarnesses and control surfaces117,601503
11Claude-MemMemory and codebase context113,896264
12Taste SkillDesign taste for agents111,134427
13Agent ReachNew capabilities109,620430
14ImpeccableDesign taste for agents104,617295
15VoiceStudioAI apps beyond coding102,675487
16ArchifyDesign taste for agents94,150459
17HyperFramesNew capabilities79,993332
18Diagram DesignDesign taste for agents65,470321
19Security AuditNew capabilities63,441450
20God's Eye ViewAI apps beyond coding63,156461
21DeepTutorAI apps beyond coding55,850178
22HallmarkDesign taste for agents49,429256
23T3 CodeHarnesses and control surfaces49,363182

Methodology and skill packs

These change how the agent works rather than where it runs. Most respond to the same complaints Andrej Karpathy listed in a widely shared post: agents make silent assumptions, overbuild and edit code they were not asked to touch.

Matt Pocock Skills mattpocock/skills

Small, composable skills that TypeScript educator Matt Pocock says he uses every day. He positions them against frameworks that own the whole process, such as GSD, BMAD and Spec-Kit, so you keep control of each step.

Check first: Choose one install method per agent. The plugin updates itself and skills.sh copies editable files; installing both loads every skill twice.

Superpowers obra/superpowers

A full development methodology for coding agents, built from composable skills plus starter instructions that make the agent actually use them. The workflow moves from brainstorming to a written plan to test-driven execution, and it supports Claude Code, Codex, Cursor and Gemini CLI.

Check first: Expect the agent to ask more questions and write plans before it codes. That pays off on multi-step features and feels slow on one-line fixes.

ECC affaan-m/ECC

An agent harness optimization system: skills, instincts and settings maintained across seven harnesses. The open-source repo is MIT-licensed. ECC Pro is a paid GitHub App for private repos.

Check first: The README warns that only the GitHub repo and its listed npm packages are official. Do not stack a manual install on top of the plugin.

Karpathy-inspired guidelines multica-ai/andrej-karpathy-skills

A single CLAUDE.md file with four principles derived from Andrej Karpathy's observations on LLM coding mistakes: wrong assumptions, bloated code and side-effect edits.

Check first: This is a community project inspired by his post, not an official Karpathy release. Merge it into your existing CLAUDE.md instead of overwriting yours.

Ponytail DietrichGebert/ponytail

A skill that pushes the agent to write only the code a task needs, modelled on the quiet senior engineer who replaces fifty lines with one. Its stated rule is to never cut validation, error handling, security or accessibility.

Check first: The project reports 53% less code and 45% fewer tokens in its own Claude Code benchmark. Run it on a slice of your own repo before you rely on those numbers.

Design taste for agents

These five repositories offer design instructions or tools for agents: art direction, imagery, accessibility and ways to steer a design. The linked Jobs and Gates video supplies historical context for the debate about software design; it does not establish these projects' quality or growth.

Video: Why Steve Jobs And Bill Gates Needed Each Other

Taste Skill Leonxlnx/taste-skill

An agent skill for building company websites with real art direction instead of a generic landing page. It uses real photographs from open archives and credits each one.

Check first: It targets marketing and company sites. For dashboards or app UI, check whether its defaults fit before you apply it everywhere.

Impeccable pbakaus/impeccable

A design vocabulary for coding agents from Paul Bakaus. You steer with commands such as polish and bolder and review with critique and audit. It works with more than 15 agents, including Claude Code, Codex, Cursor, Grok Build and DeepSeek Harness.

Check first: Run its init step inside each project, because the steering depends on project context.

Archify tt-a1i/archify

Turns an idea, a plan or a whole codebase into an interactive architecture diagram with links back to the source. It works with Cursor, Claude Code, Codex CLI and OpenCode.

Check first: The suggested install asks your agent to read a remote llms.txt file. Read any install prompt yourself before your agent runs it.

Diagram Design cathrynlavery/diagram-design

An agent skill that draws diagrams as one self-contained HTML and SVG file in your brand's colours and fonts. It has 45 diagram types and imports from draw.io, Mermaid and Excalidraw.

Check first: Run its brand onboarding against your site once so later diagrams match your styling.

Hallmark Nutlope/hallmark

A design skill from Together AI that "refuses to look AI-generated." It picks from 50 material-based themes and checks the result in a real browser against hard gates: contrast, focus states, reduced motion, invented facts and broken phone layouts.

Check first: Use its audit command first on an existing page. It only reports and does not rewrite anything.

Harnesses and control surfaces

Harnesses decide where the agent runs, which models it can call and how you supervise it. DeepSeek now ships its own. T3 Code and Orca sit one level up and drive several agents at once. T3 Code's supported list already includes xAI's Grok Build CLI.

DeepSeek Harness deepseek-ai/deepseek-harness

DeepSeek's own open-source agent harness (dsh). It is built on an everything-is-a-plugin architecture powered by the Cordis framework and runs in the browser with one npx command.

Check first: It is a developer preview, and the README warns of compatibility-breaking changes. Read its SAFETY.md before pointing it at real work.

OpenCode anomalyco/opencode

An open-source coding agent installable through curl, npm, Homebrew, Arch packages, mise or Nix. Many skills on this page list it as a supported host.

Check first: There are two Homebrew formulas. The project recommends its own tap; the official formula updates less often.

Orca stablyai/orca

Runs Claude Code, Codex and other agents in parallel, each in its own git worktree. One agent can split a large task across workers and merge the results, on your laptop, an SSH host or a cloud VM.

Check first: Parallel workers multiply token spend. Decide on a budget per task before you hand over large jobs.

T3 Code pingdotgg/t3code

An "agent harness control surface" that drives the agents already on your machine, including Claude Code, Codex, Cursor, OpenCode, Antigravity and Grok Build, with iOS and Android apps for remote control.

Check first: It uses your existing subscriptions. Install and sign in to at least one provider's CLI before you start it.

Memory and codebase context

Both tools here attack the same cost: an agent re-reading the same files and re-learning the same project every session.

Graphify Graphify-Labs/graphify

Maps a project's code, docs, PDFs, images and video into a knowledge graph that the agent queries instead of grepping. Code is parsed locally with tree-sitter, with no LLM involved. A paid Graphify Cloud tier adds always-on graphs.

Check first: Code stays on your machine, but docs, PDFs, images and video are sent to your model provider. Exclude sensitive files.

Claude-Mem thedotmack/claude-mem

Persistent memory across sessions. It records what your agent does, compresses it with AI and injects the relevant context into later sessions. It works with Claude Code, Codex, Gemini, Copilot, OpenCode and others.

Check first: It stores observations from every session. If you handle credentials or customer data, review what it captures and where it is stored.

New capabilities

These give an agent skills it does not have out of the box: reverse engineering, reading platforms it cannot reach, structured security audits and rendering video.

REA morluto/rea

"Reverse Engineer Anything": one MCP server for reverse engineering binaries, applications and runtime behavior, so you can trace how a feature works down to the binary level.

Check first: The README warns that look-alike repos and copied READMEs are not official. Install only from morluto/rea, rea.tools or the rea-agents npm package, and only analyze software you are allowed to.

Agent Reach Panniantong/Agent-Reach

Gives an agent access to platforms it cannot reach on its own, including searching and reading X, Reddit, YouTube and Xiaohongshu.

Check first: Automated reading of some platforms breaks their terms of service. Check each platform's rules and prefer official APIs where they exist.

HyperFrames heygen-com/hyperframes

HeyGen's open-source framework for rendering HTML, CSS, media and animation into deterministic MP4 video. It runs from a CLI, from coding agents via skills, or in a free desktop studio.

Check first: The framework and studio are free; the Pro plan adds HeyGen's paid generation models. Know which one a template relies on.

Security Audit cloudflare/security-audit-skill

Cloudflare's coding-agent skill for multi-phase security audits: reconnaissance, coverage-led hunting, validation by fresh verifier agents and a structured findings report. It seeded Cloudflare's own vulnerability discovery harness.

Check first: Treat findings marked needs_validation as leads, not confirmed bugs. Only confirmed findings come with a full source trace.

AI apps beyond coding

These end-user AI applications broaden the comparison beyond agent add-ons. Inclusion does not establish that their growth matches the other projects.

VoiceStudio debpalash/VoiceStudio

A fully local, open-source alternative to ElevenLabs for voice cloning, video dubbing, voice design, dictation and narration. The README lists 646 languages.

Check first: Running locally keeps audio private, but cloning still needs consent from the person whose voice you use.

God's Eye View bilawalsidhu/gods-eye-view

A photorealistic 3D globe in the browser that shows live public data, including aircraft, ships, satellites, earthquakes, traffic and public cameras, with voice control powered by a realtime AI agent.

Check first: It starts without API keys. Optional keys add paid data sources, and the README has a section on keys and costs.

DeepTutor HKUDS/DeepTutor

A personalized AI tutoring workspace from the University of Hong Kong's data intelligence lab, with learning paths, notebooks and narrated video lessons. It ships releases about every week.

Check first: Its fast release pace means breaking changes are likely. Pin a version if you build on it.

Before you install any of these

Skills and plugins are instructions your agent follows with your permissions. Install from the official repository, not a fork or a copied README. Read the skill files before enabling them, and pin a version for anything that runs in CI. Check which files are sent to a model provider, and give agents only the credentials a task needs.

Sources

  • Star counts, creation dates and licenses: GitHub REST API, fetched November 6, 2026.
  • Descriptions and check notes: each project's README on GitHub, linked from its heading above.

Projects are not presented as personally tested unless marked "From my own use." Corrections: contact.