AI Tools

Best GitHub Repositories for AI Development in 2026

Néstor Vázquez
Néstor Vázquez
July 24, 2026 · 20 min read

Building with AI today means more than prompting a chatbot — it means wiring together skills, agents, and tools that plug directly into your coding workflow. The best of that ecosystem lives on GitHub, usually buried under thousands of forks, abandoned experiments, and repos that peaked at a README.

This is a running list of the ones actually worth starring: real tools solving real problems for developers working with Claude Code, Cursor, Copilot, Windsurf, and the rest of the AI coding stack. We'll keep adding to it. Here's where it starts.

1. UI UX Pro Max — Design Intelligence for AI Coding Agents

Repo: nextlevelbuilder/ui-ux-pro-max-skill · MIT licensed · 100k+ stars

UI UX Pro Max is an AI skill — installable straight into Claude Code, Cursor, Copilot, Windsurf, and a dozen other agentic coding tools — that gives those agents actual design judgment instead of generic, interchangeable-looking output. Point it at a brief like "build a landing page for my beauty spa," and its reasoning engine cross-references 161 industry-specific rules, 67 UI styles, 161 color palettes, and 57 font pairings to return a complete, justified design system: layout pattern, color palette, typography pairing, motion rules, and a checklist of anti-patterns to avoid for that specific industry. It ships as both an MIT-licensed CLI (ui-ux-pro-max-cli) and a native Claude Code skill, and within months of launch it passed 100,000 GitHub stars — a sign of just how many people are trying to solve "make it stop looking AI-generated" with the same tool.

2. OpenCLI — Turn Any Website Into a CLI for AI Agents

Repo: jackwener/OpenCLI · Apache-2.0 licensed · 27k+ stars

OpenCLI gives AI coding agents a deterministic way to operate the web instead of guessing at DOM selectors on every run. Install the opencli-browser skill in Claude Code, Cursor, or any other agentic tool, and it can navigate, click, fill forms, and extract data through your actual logged-in Chrome session — no separate scraping auth, no headless browser fighting login walls. It also ships built-in adapters for sites like HackerNews, Reddit, Bilibili, and Twitter/X, works as a CLI hub for tools like gh and docker, and lets you write new site adapters end-to-end with a guided skill. Apache-2.0 licensed, and closing in on 30k stars less than five months after its first commit.

3. Engram — Persistent Memory for AI Coding Agents

Repo: Gentleman-Programming/engram · MIT licensed · 5.6k+ stars

Engram solves the problem every agentic coding session eventually hits: the agent forgets everything the moment the session ends — every architecture decision, every bug fix, every pattern it discovered, gone. Engram is a single Go binary (SQLite + FTS5 under the hood) that gives any MCP-compatible agent — Claude Code, Cursor, Windsurf, Gemini CLI, Codex, VS Code Copilot, and more — a persistent, searchable memory that survives across sessions. No Node service, no Python stack, no hosted account required for the core local path; an opt-in cloud tier adds project-scoped sync and a dashboard on top. Install via brew install gentleman-programming/tap/engram, run engram setup <your-agent>, and it wires itself into your agent's MCP config automatically.

4. Awesome AutoResearch — The "Leave an Agent Running Overnight" Playbook

Repo: WecoAI/awesome-autoresearch · CC0 licensed · 1k+ stars

AutoResearch started as a single markdown file from Andrej Karpathy: point a coding agent at one file, give it an evaluation metric, and let it loop — edit, run, check if the metric improved, keep or revert, repeat. Applied to nanoGPT training code, that loop found real speedups overnight on hardware that would otherwise sit idle. Awesome AutoResearch is the curated list tracking what happened next: the pattern getting adapted to GPU kernel tuning, template-engine optimization, tabular ML, voice-agent prompt tuning, and more — each entry linked to the actual optimization trace, not just a claimed result, so you can see exactly what the agent tried before it landed on the win. If you've been wondering whether "give an agent a metric and a time box" generalizes beyond ML training, this is the running evidence.

5. pi-autoresearch — The AutoResearch Loop, Packaged as a Reusable Extension

Repo: davebcn87/pi-autoresearch · MIT licensed · 7.2k+ stars

Where Awesome AutoResearch tracks one-off use cases, pi-autoresearch turns the pattern into an installable tool for pi, the terminal-based coding agent. Run pi install npm:pi-autoresearch and you get an /autoresearch command plus a skill that handles the whole loop for you: it asks what to optimize and how to measure it, then repeatedly tries an idea, benchmarks it, keeps the change if the metric improved, reverts it if not, and auto-commits along the way — with a live dashboard you can open in the browser to watch it work. It's not tied to ML training; people have pointed it at test-suite speed, bundle size, build times, and Lighthouse scores. Shopify CEO Tobi Lütke used it on the company's Liquid template engine and shipped a 53% faster parse+render time from 93 fully automated commits — a solid demonstration that "leave an agent running against a metric overnight" works on production codebases, not just toy benchmarks.

6. deepsec — An Agent-Powered Vulnerability Scanner From Vercel

Repo: vercel-labs/deepsec · Apache-2.0 licensed · 6.2k+ stars

deepsec points coding agents at your codebase specifically to find security bugs — not the pattern-matching of a typical static analyzer, but an actual investigation: it starts with static analysis to flag security-sensitive files, then hands each candidate to an agent (Claude or Codex, run at maximum thinking effort) to trace the data flow, check for existing mitigations, and produce a scored finding instead of a guess. It's built to run on your own infrastructure — nothing about your codebase gets sent to a third party — and fans work out across worker machines for large repos, resuming cleanly if a run gets interrupted. Vercel built it for internal use first, then open-sourced it after running it against major OSS projects and getting real signal out of it. It isn't cheap to run at full strength — scans on large codebases can cost thousands of dollars given the reasoning effort involved — but for teams sitting on vulnerabilities that have gone unnoticed for years, that's the trade being made.

7. LeanCTX — A Context-Compression Layer That Cuts Agent Token Spend 60–90%

Repo: yvgude/lean-ctx · Apache-2.0 licensed · 3.3k+ stars

LeanCTX sits between your coding agent and everything it touches — files, shell output, the model itself — and decides what actually needs to go into the context window. A repeated file read that would normally cost ~2,000 tokens gets served from cache for around 13; a raw git status dump that would run ~800 tokens gets compressed to roughly 120. It also adds an optional local proxy that shrinks every outgoing request (system prompt, history, tool results) in a prompt-cache-safe way, persists session memory across chats instead of resetting each time, and ships a real-time dashboard so you can actually see where your tokens are going instead of guessing. It's a single local Rust binary, works with 30+ agents including Claude Code, Cursor, Copilot, Windsurf, Codex, and Gemini, and needs zero config to start saving tokens.

8. Traffic Source — Self-Hosted Web Analytics With Conversion Tracking Built In

Repo: mddanishyusuf/traffic-source · no license specified · 613+ stars

Traffic Source is a self-hosted alternative to Google Analytics that goes further than most privacy-first trackers by wiring in the parts that actually tie traffic to revenue: real-time pageviews, referrers, and UTM breakdowns, geo and device data via Cloudflare proxy headers, a one-click Google Search Console connection you can drill into by keyword, and a Stripe integration that polls for payments every 60 seconds — no webhooks to configure. It also ships a built-in affiliate system with custom commission rates and shareable referral links, plus full session replay of a visitor's path before they converted. No cookies, a ~3KB tracking script, and SQLite under the hood, so there's no external database to stand up — deploy it on a $4/month VPS, or one-click it straight onto Railway.

9. SEO Machine — A Claude Code Workspace for Content That Actually Ranks

Repo: TheCraigHewitt/seomachine · MIT licensed · 7.2k+ stars

SEO Machine turns Claude Code into a full SEO content pipeline instead of a general-purpose writing assistant. Slash commands like /research, /write, /analyze-existing, /optimize, and /publish-draft hand off to specialized agents for content analysis, meta elements, internal linking, keyword mapping, editing, and headline generation — with Google Analytics 4, Search Console, and DataForSEO wired in for real performance data instead of guesswork. The part that keeps it from producing generic AI-blog-slop is the context layer: you fill in your own brand voice, style guide, target keywords, competitor analysis, and a handful of exemplary posts before it writes a word, and every draft gets scored 0–100 against readability, keyword density, and search-intent match. Craig Hewitt built it for his own company, Castos, before open-sourcing it — the repo ships a complete real-world example of every context file filled out so you're not starting from a blank template.

10. Scrapling — Web Scraping That Survives the Site Changing Under It

Repo: D4Vinci/Scrapling · BSD-3-Clause licensed · 71k+ stars

The thing that breaks most scrapers isn't the anti-bot wall, it's the site quietly redesigning itself six months later — every hardcoded selector dies at once. Scrapling attacks that problem directly: its parser learns the shape of the elements you scraped and can relocate them automatically after a layout change, so a script you wrote once keeps working instead of needing constant babysitting. On top of that it ships fetchers that get past anti-bot systems like Cloudflare Turnstile out of the box, and a full async spider framework — Scrapy-like API, concurrent crawling with per-domain throttling, mixing plain HTTP with headless and stealth browser sessions in the same run — for when you need to go from "grab one page" to "crawl the whole site" without switching libraries. It also ships an MCP server and an agent skill, so coding agents can drive it directly instead of you writing the scraping code by hand.

11. SynthID Text — Google DeepMind's Reference Implementation for Watermarking LLM Output

Repo: google-deepmind/synthid-text · Apache-2.0 licensed · 930+ stars

A change of pace from the coding tools above, but relevant to anyone building on top of LLMs: SynthID Text is Google DeepMind's reference implementation for embedding an invisible statistical watermark into text generated by a language model, and detecting it afterward. Instead of altering the output's meaning, it nudges the token-selection probabilities during generation using a pseudo-random function — imperceptible to a human reader, but recoverable by a detector even after the text has been generated. The repo ships as a Hugging Face Transformers mix-in for Gemma and GPT-2, plus both a no-training-required Weighted Mean detector and a more powerful trainable Bayesian detector. It's explicitly a research/reproducibility reference rather than a production library — Google points production use toward the built-in SynthID support already shipped in Hugging Face Transformers — but it's the closest thing to primary-source code behind the technique now used across Google's own AI products to flag AI-generated content, and the underlying method was peer-reviewed and published in Nature.

12. docker-android — A Disposable Android Emulator for CI and Agent Testing

Repo: HQarroum/docker-android · MIT licensed · 7k+ stars

docker-android packages a fully functional, remotely controllable Android emulator into a Docker image — Alpine-based, KVM-accelerated, with ADB and the emulator's network ports forwarded straight to the container. Point docker compose up at it and you get a clean Android device that wipes itself on every restart, which is exactly what you want for CI pipelines running app tests, or for an AI agent that needs a disposable device to drive through a mobile app end-to-end without touching a real phone. It's configurable down to the API level, device type, and image variant — you can build a lean 138MB image without the SDK and emulator for stripped-down use cases, or the full ~2GB variant with Google Play Store baked in — and it plays nicely with scrcpy for remotely watching the emulator's screen while it runs headless.

13. /last30days — A Search Skill That Scores by Engagement, Not Editors

Repo: mvanhorn/last30days-skill · MIT licensed · 53k+ stars

/last30days is an agent skill built on a simple observation: Google indexes what editors wrote, but Reddit upvotes, X likes, YouTube engagement, and Polymarket odds reflect what real people are actually paying attention to — and no single search engine touches all of those signals at once because each platform is its own walled garden. This skill bridges them: point it at a topic or a person, and it searches Reddit, X, YouTube, HackerNews, Polymarket, GitHub, TikTok, Instagram, and Bluesky in parallel, then has an agent score and synthesize the results into one grounded brief instead of dumping ten separate feeds on you. Reddit, HackerNews, Polymarket, and GitHub work with zero setup; a 30-second wizard unlocks the rest once you're ready to bring your own keys. It installs straight into Claude Code via the plugin marketplace, or into 50+ other agent hosts through npx skills add, and it hit #1 on GitHub Trending within weeks of release.

14. TimesFM — A Foundation Model for Forecasting, Zero-Shot

Repo: google-research/timesfm · Apache-2.0 licensed · 27k+ stars

Another one outside the coding-agent lane, but worth knowing about if your work touches forecasting: TimesFM is Google Research's pretrained foundation model for time-series data — pretrained on 100 billion real-world and synthetic time points, it can forecast an entirely new series zero-shot, no fine-tuning or task-specific training required, the same way an LLM can handle a new prompt it's never seen. The latest version (2.5) runs at 200M parameters with up to 16k of context and supports probabilistic forecasts out to a 1k-step horizon. It's already wired into Google's own stack — BigQuery ML for SQL-based forecasting at enterprise scale, Connected Sheets for spreadsheet forecasting, and a Dockerized endpoint in Vertex Model Garden built specifically for agentic calling — and the repo itself now ships an AGENTS.md and a SKILL.md, so a coding agent can pick it up and use it directly rather than you writing the integration by hand.

15. GEO SEO SuperApp — A Free macOS Workspace for SEO and GEO Agencies

Repo: on-page-ai/geo-seo-superapp · macOS desktop app

A different kind of entry — this one's brand new, not an open-source library. GEO SEO SuperApp is a free native macOS app from On-Page.ai that bundles the recurring parts of agency SEO work into one workspace: page reports and guided optimizations from a keyword and URL, technical site crawls with findings routed into deeper analysis, Google Search Console and Bing Webmaster Tools review, scheduled scans with visible browser QA, indexation and sitemap tracking, and reusable "recipes" you can build and share across clients. It signs in with your existing ChatGPT account for the AI agent side and connects to an On-Page.ai account for the core SEO data — both required, everything else (GSC, Bing, Cloudflare, browsers, social accounts, MCP services) is optional and added later through the Setup Center. The GitHub repo is really a signed, notarized release channel for the app rather than a source-available library — worth knowing if you were expecting to read or fork the code.

16. mcp-cli — A Shell-Native Way to Call MCP Servers Without the Context Bloat

Repo: philschmid/mcp-cli · MIT licensed · 1.2k+ stars

Every MCP server you connect to an agent dumps its full tool schema into context up front, whether you use it or not — that adds up fast once you have more than a couple wired in. mcp-cli, built by Philipp Schmid (DevRel lead at Google DeepMind), sidesteps that by turning MCP servers into something you call from the shell instead: mcp-cli lists servers and tools, info shows a tool's schema on demand, and call actually invokes it, with JSON in and out so it pipes cleanly into jq and the rest of your scripts. A lazy-spawned daemon keeps connections warm between calls, it supports both stdio and HTTP MCP servers, and it compiles to a single standalone binary via Bun — no runtime to install. It's built specifically with AI coding agents in mind (Gemini CLI, Claude Code, and friends), giving them the ability to discover and call any MCP tool on demand rather than paying the token cost of every tool definition on every turn.

17. MCPorter — Turn Any MCP Server Into Typed TypeScript or a Standalone CLI

Repo: openclaw/mcporter (originally steipete/mcporter) · MIT licensed · 4.8k+ stars

MCPorter leans into the "code execution" pattern Anthropic has been pushing for MCP: instead of an agent burning context re-discovering a server's schema on every call, MCPorter discovers what's already configured on your system — pulling from Cursor, Claude, Codex, Windsurf, OpenCode, and VS Code's own configs — and lets you call it directly, no boilerplate. Point mcporter generate-cli at any MCP server and it mints a standalone, compilable CLI out of it in one command; emit-ts instead generates typed TypeScript client wrappers so an agent or test suite gets real autocomplete and validation instead of hand-rolled JSON-RPC plumbing. It also ships record/replay for capturing MCP traffic as fixtures, built-in OAuth caching, and support for HTTP, SSE, and stdio transports through one consistent API. Built by Peter Steinberger and now maintained under the OpenClaw org.

18. Impeccable — Design Guidance So Your AI-Built UI Stops Looking AI-Built

Repo: pbakaus/impeccable · Apache-2.0 licensed · 49k+ stars

Every model got trained on the same SaaS templates, so skip design guidance entirely and you get the same tells on every AI-built interface: Inter everywhere, purple-to-blue gradients, cards nested in cards, gray text on colored backgrounds, a rounded-square icon tile above every heading. Impeccable started from Anthropic's frontend-design skill and built it out into a full shared vocabulary between you and your AI coding tool: one /impeccable init writes a PRODUCT.md and DESIGN.md so every later command knows your audience, brand lane, voice, and component conventions, then 23 commands — polish, critique, bolder, quieter, animate, harden, and more — give you a precise word for the kind of change you want instead of vague prompting. Underneath, 60 deterministic detector rules run without an LLM or API key at all, catching accessibility, performance, and responsive issues the same way every time. Creator Paul Bakaus has kept iterating in public — v3.5 moved from designing in isolation to iterating on real UI directly in the codebase you ship.

19. OpenTUI — A Native Zig Core for Building Terminal UIs in React, Solid, or Plain TypeScript

Repo: anomalyco/opentui · MIT licensed · 12.7k+ stars

Terminal UIs used to mean picking a framework in whatever language you were already writing your CLI in and living with its limits. OpenTUI decouples the two: the rendering core is written in Zig and exposed through a C ABI, so it's usable from any language, while the TypeScript bindings on top give you actual component-based UI development — React and SolidJS reconcilers included — with flexible layout instead of manually pushing characters around a grid. It's built by Anomaly, the team behind OpenCode, specifically because they hit the ceiling of their previous Go/Bubble Tea terminal UI and needed something faster and more composable; OpenTUI now powers OpenCode in production and will power terminal.shop next. There's also an installable AI agent skill (npx skills add anomalyco/opentui --skill opentui) that teaches your coding assistant OpenTUI's APIs directly, so it can build terminal interfaces with the right patterns instead of guessing.

← Back to all articles

Keep Reading

19 SEO Experts I Follow in 2026

A working list of the SEO professionals whose work actually changes how I audit, implement, and mentor clients in 2026.