AI page
Repos people are talking about
Ranked by growth inside the window: stars and forks gained, percentage growth, mentions, and how recently the code was pushed. Size alone counts for nothing. Each repo read and explained by AI.
- 1
Hindsight is a Python-based agent memory system that aims to let AI agents learn over time rather than just recall past conversation history. It exposes retain/recall/reflect operations, memory banks, mental models, and an MCP server, runs as an API server or embedded in Python, and claims state-of-the-art accuracy on the LongMemEval benchmark, outperforming RAG- and knowledge-graph-based approaches.
Why now: It was a GitHub Trending repository with rapid star growth and published benchmarks reportedly independently reproduced by Virginia Tech and The Washington Post, fueling interest.
- 2
Paperclip is an open-source orchestration platform for running teams of AI agents like a company. It's a Node.js server with a React UI where you define business goals, 'hire' agents (OpenClaw, Claude Code, Codex, Cursor, bash, HTTP), assign roles and budgets, then monitor work and costs from a dashboard. It provides org charts, governance, goal alignment, and coordination for autonomous agent organizations.
Why now: It has exploded in popularity with 86k+ stars, riding the wave of interest in coordinating many autonomous AI agents, and positions itself alongside trendy tools like OpenClaw and Claude Code.
- 3
Laya is a Python library that makes fast, single-forward-pass decisions over text — yes/no answers, scores, or choices from typed options — in 100+ languages. Instead of generating tokens like an LLM, it uses non-autoregressive checkpoints (ModernBERT-based) trained with RL, plus a router that automatically picks the right model per request, running in tens of milliseconds.
Why now: It's trending as a newly released, heavily starred project, amplified by a dev.to article where the author says a frontier lab validated his earlier non-autoregressive decision-model idea.
- 4
Jev Ultrafast is a Python browser agent that completes a natural-language goal by controlling a real Chrome browser. Instead of a large model choosing free-form actions, it builds an indexed table of page elements and uses one TypeSafe API call to pick an operation (click, select, scroll, wait) and target, with a small LLM only generating text for typing. The README demos a Google Flights search in ~7 seconds.
Why now: It recently appeared on GitHub's 'new repos, most starred' lists with over 20,000 stars in days, driven by a viral speed demo and the opening of the Browser Use Cloud waitlist.
- 5
ZCode is an AI coding workbench from Z.ai that ships as a desktop app, browser-based interface, and terminal agent. The monorepo includes the client, backend services, shared UI, and the source for its Agent CLI and runtime, which can run a TUI or local web UI via a single `zcode` command, with optional SSH/WSL remote workspace support.
Why now: Recently released v3.14.3 and quickly climbing GitHub trending with ~6.8k stars within days of the repo's creation.
- 6
Jev 聊天助手 is an Android 'conversation copilot' that reads the chat currently on your screen (via accessibility services, no app hooking or repackaging), uses LLMs to analyze the other person's intent and draft candidate replies, then fills them into the input box — you decide whether to send. It supports QQ, X DMs, and Feishu (with OCR fallback), and lets you configure your own model APIs.
Why now: It's trending on GitHub's most-starred new repos (~6.7k stars days after its September 2026 release), notable for its non-invasive screen-reading approach and its decision-then-draft pipeline.
- 7
Kev is a family of small open decision models built on Qwen3.5/3.8, following the Jev architecture. They answer yes/no, multiple-choice, and rating questions with calibrated probabilities. You can use pretrained weights in four sizes (0.8B to 27B), fine-tune on your own labeled data, and serve them locally or with a one-command deploy that scales to zero.
Why now: It's an open, self-hostable drop-in replacement for TypeSafe's hosted Jev API — the TypeSafe SDK works unchanged — and benchmarks show the 27B model approaching Jev's accuracy.
- 8
Univer is an open-source TypeScript SDK for embedding office-style apps—spreadsheets, docs, presentations, relational tables, and more—into your own product. It offers canvas-based rendering, a formula engine, a plugin architecture, and a Facade API that works in both the browser and Node.js, letting developers compose just the features they need rather than adopting a fixed UI or hosted app.
Why now: It is repositioning itself as 'The Office Harness for AI Agents,' a timely pitch as teams look for embeddable productivity surfaces where AI agents and humans can work on the same files.
- 9
Laya-MLX is a native Apple Silicon (MLX) runtime for Laya typed decision models: instead of generating text, it returns structured answers (e.g. picking a department from a choice list) for short English questions. It runs fully locally with no PyTorch, Transformers, or cloud API, achieving ~13 ms median latency (7.4 ms multilingual) on an M3 Max, with a pip-installable Python API and a Snake-game demo.
Why now: It's trending among newly starred repos, likely due to the low-latency local typed-decision demos (Snake at 75 moves/s) and benchmark claims versus text-generation approaches.
- 10
A free, open-source MIT-licensed curriculum of 523 lessons across 20 phases (~342 hours) teaching AI engineering from scratch — math, machine learning, LLM apps, agents, MCP servers, computer vision — in Python, TypeScript, Rust, and Julia. Each lesson produces a reusable artifact like a prompt, skill, agent, or MCP server, with a companion website and multilingual translations.
Why now: It has rapidly grown to 58k+ stars and heavy traffic; its popularity likely comes from demand for practical, hands-on AI engineering education amid a widely cited gap between AI tool usage and professional readiness.
- 11
Compositor is a free, open-source image editor for Mac built in Swift as a Photoshop alternative. It focuses on compositing and post-processing workflows, offering layers, masks, adjustment layers, selections, painting and retouching tools, filters, and a Camera Raw-style panel, with the full Xcode project available so anyone can modify features.
Why now: It has been rapidly gaining stars on GitHub as a trending new project, tapping into frustration with Adobe's pricing and promising a familiar, genuinely full-featured free alternative on Mac.
- 12
CLM is a Python library serving Contrastive Language Model-8B, a model trained contrastively to connect states and actions for fast 'System One' decision-making. It exposes a TypeSafe-compatible API for scoring states, asking typed questions (boolean/choice/score), and ranking candidate answers. A vLLM embedding server plus a small 75 MB reference head make serving lightweight, with cached state/action embeddings for low latency.
Why now: It just launched as a highly starred new repo (v0.1, ~1.5k stars within a day), claiming performance on par with a rival 'Jev' at up to 9× lower latency and new SOTA verifier results on Terminal-Bench 2.1 and DeepSWE, plus an open invitation to benchmark it.
- 13
NVIDIA Model Optimizer (ModelOpt) is a Python library that compresses deep learning models using techniques like quantization, pruning, distillation, sparsity, neural architecture search, and speculative decoding. It takes Hugging Face, PyTorch, or ONNX models, applies optimization via Python APIs, and exports optimized checkpoints ready for deployment in inference frameworks like TensorRT-LLM, TensorRT, vLLM, and SGLang for faster inference.
Why now: Recent NVFP4 quantization news and blogs, including Nemotron 3 Ultra 550B NVFP4 checkpoints and W4A4 tutorials, plus new AutoQuantize and quantization-aware distillation posts.
- 14
A set of Claude Code skills that turn a chess game into a readable post-mortem: it drives Stockfish to analyze every move, generates plain-language explanations, an annotated PGN, an HTML analysis board, and a narrated video. You can also supply think-aloud audio or notes, which the tool transcribes (whisper.cpp) and matches to moves so it critiques your actual reasoning.
Why now: A Show HN post demonstrating the narrated-video output and the novel idea of Claude driving Stockfish to answer human-style 'why' questions
- 15
apk-reverse is an Agent Skill for Android APK reverse engineering. Rather than a tutorial, it's a structured procedure meant to be loaded by AI coding agents (Claude Code, Codex, etc.), guiding them through debloating, ad removal, dex patching, repacking, and runtime/server analysis using override rules, gates, symptom indexes, and helper scripts to prevent agents from making repeated mistakes.
Why now: It's trending as a newly created, heavily starred repository, notable for its novel approach of encoding reverse engineering workflows as agent-consumable skills with guardrails against typical LLM failure modes.
- 16
SemIf is an open-source Python project that lets software make small semantic decisions (routing, retries, yes/no checks) using open LLMs like a 4B Qwen model, reading typed option probabilities directly instead of generating and parsing text. It reproduces the interface pattern of TypeSafe's closed Jev service and runs locally on a consumer GPU, Apple Silicon, or even CPU, with a browser demo.
Why now: It's trending as a popular open-source alternative to TypeSafe's closed Jev service — recently renamed from OpenJev, with active development, a no-waitlist browser demo, and 4.3k stars within a week of creation.
- 17
Buzz is a self-hostable team workspace where humans and AI agents collaborate in shared channels. Built in Rust on a Nostr-style relay, every message, patch, review, and workflow step is a signed event in one log, so agents act as members with their own keys and the same capabilities and audit trail as people.
Why now: It rides the current wave of interest in agentic development tools, standing out by giving agents full workspace participation under a self-hosted, signed-event model rather than bot-style integrations.
- 18
A Claude Code plugin and npm library that replaces lossy LLM-based context compaction with verbatim pruning: instead of summarizing old turns, it scores every tool call and result in fast API requests and deletes or truncates only the stale ones. Kept user and assistant text stays verbatim, recent messages are pinned, and it can also be used as a standalone TypeScript package.
Why now: It appeared among the most-starred new repos, likely because it offers a fresh alternative to lossy context-window compaction in Claude Code workflows.
- 19
AirCard is a macOS app that lets you customize the appearance of Apple Wallet cards and the iPhone lock screen passcode keypad on iOS 18+ without jailbreaking. It detects cards via a connected iPhone, lets you assign custom artwork per card or in bulk, and includes a creator for .passthm passcode themes with photo framing and real-time previews. It works standalone with no extra tools required.
Why now: It's gaining attention as one of the most-starred new repos, riding interest in no-jailbreak iOS theming via the 'airlift' exploit.
- 20
Unreal Agent is a Go library for building LLM agents around an async-first architecture. It provides durable, idempotent input handling, append-only forkable session history, tool translators that validate model calls without I/O, and an actor-based operation manager for durable asynchronous work. Components are interface-based and swappable, e.g. proxying operations to remote sandboxes.
Why now: It recently appeared as one of the most-starred new repos, gaining ~2k stars within days of release.
- 21
Search is a lightweight, free macOS web browser built on WebKit, made by design studio Office Commun. It strips browsing down to tabs and a single address/search field, with built-in ad blocking, reading mode, element hiding, picture-in-picture video, keychain password storage, and support for Chrome extensions via WebKit's extension engine. No accounts, sync, or telemetry; about 3 MB in size.
Why now: It was featured on Hacker News, likely drawing attention as a minimalist alternative to bloated browsers like Chrome and Arc, with Chrome extension support on WebKit as a notable hook.
- 22
reverse-skill is a skill router pack for AI coding agents (Claude Code, Cursor, Codex, Cline) that handles reverse engineering and authorized penetration testing tasks. When given an APK, binary, JS encryption, PCAP, or CTF challenge, it routes the agent to the right methodology and playbook, checks for required tools like jadx, Frida, or BurpSuite, and enforces a repeatable evidence-driven workflow with reports and a self-evolving knowledge base.
Why now: It is trending (Trendshift badge, ~37.8k stars) as AI agents increasingly take on security/reverse-engineering work; a v1.0.1 release exists but no specific recent event is clear from the sources.
- 23
OpenBao is an open-source tool for managing, storing, and distributing sensitive data like secrets, certificates, and keys. It encrypts secrets at rest, generates dynamic on-demand credentials (e.g., AWS keys, database logins) with automatic revocation, offers encryption-as-a-service, and provides leasing, renewal, and audit logging under a community-governed open-source license.
Why now: It is the community-driven open-source fork of HashiCorp Vault created after Vault's license change, drawing attention from teams seeking an OSI-approved, open-governance alternative.
- 24
Next.js is Vercel's React framework for building full-stack web applications. It extends React with features like server-side rendering, static site generation, and hybrid rendering, and integrates Rust-based JavaScript tooling for fast builds. It's widely adopted, from individual developers to some of the world's largest companies.
Why now: not clear from the sources
- 25
Visual Studio Code is Microsoft's open-source code editor (the Code - OSS repository), combining simple editing with debugging, code navigation, an extension system, and integration with existing tools. Built in TypeScript on Electron, it ships monthly releases for Windows, macOS, and Linux, and is developed openly with the community under the MIT license.
Why now: It remains one of the most-starred and widely used editors on GitHub, with ongoing monthly feature updates; no specific new event is clear from the sources.
- 26
OpenMuse is an open-source personal AI agent template built with CopilotKit and AG-UI for iOS, Android, and web. It gives the agent its own computer—a persistent Chromium browser, optional Linux terminal, and file workspace—so you can ask for an outcome, review its plan and actions, and pause, resume, or take over the browser session yourself.
Why now: Freshly released and gaining traction on GitHub's most-starred new repos, backed by 38-second mobile and web demos showing the agent browsing and handling email.
- 27
Bespoke Nimble is an open recipe for building a fast, typed classification model. Given text and a schema of questions (multiple-choice or true/false), Bespoke-Nimble-9B returns each answer with per-choice probabilities in a single step, no chain-of-thought reasoning. The repo covers contrastive data curation, LoRA fine-tuning of Qwen3.5-9B, and serving, runnable on Apple Silicon or NVIDIA GPUs.
Why now: It's trending as a new, highly starred open replication of TypeSafe's Jev System One model, with a recently released 9B checkpoint, public training data, and a shared benchmark suite.
- 28
Jianying Headless is a local automation tool for Jianying Pro (剪映) on macOS that turns structured JSON editing plans into editable drafts with multi-track video, audio, text, keyframes, and effects. It can modify existing projects in isolated copies and export verified H.264/AAC MP4s via the native Jianying engine, with a bundled Agent Skill and CLI. A Windows FFmpeg export path exists but doesn't produce editable drafts.
Why now: It's gaining stars as a trending new repo, showcasing a validated 23-track Hypit-to-Jianying project handoff that passed full 1507-frame native export checks — notable for AI-driven video editing workflows.
- 29
TensorFlow is Google's open-source end-to-end machine learning framework. It lets researchers train neural networks and lets developers build and deploy ML-powered applications, with a broad ecosystem of tools and libraries. It provides stable Python and C++ APIs, other language bindings, GPU acceleration via CUDA, and pip-installable packages including a lighter CPU-only build.
Why now: not clear from the sources
- 30
Whiteboard is an open-source desktop app where developers and AI coding agents collaborate on software design. Agents like Claude Code or Codex plug in via an SDK and draw diagrams (sequence diagrams, ER diagrams) on a shared canvas, with clickable links from visuals to code, an AST-aware semantic diff viewer written in Rust, and a decision log showing what the agent chose autonomously.
Why now: It was just featured on Hacker News as a Show HN launch from a YC W26 company, which is driving attention.
- 31
awesome-jev is a curated list of public projects, integrations, and discussions that use Jev — TypeSafe AI's 'System One' model that turns unstructured state plus a typed question into a typed decision (choice, score, or boolean) with a confidence value. The list organizes real-world usage by category and application domain, with strict inclusion rules and explicit warnings that listing is not endorsement.
Why now: It appeared among the most-starred new repos, likely riding interest in the Jev launch and its typed-decision approach; it also explicitly addresses controversy around bulk, low-quality submissions.
- 32
NanoJev is a small open-source Python replica of Jev, a 0.6B 'parallel decision' model. Instead of decoding tokens, it takes game states and questions and directly outputs probability distributions over 2–255 candidate choices. The repo includes an end-to-end training pipeline, datasets, and a single checkpoint that plays ViZDoom, Maze, and Snake.
Why now: A recent update (Sept 20, 2026) showed one checkpoint beating the original Jev on ViZDoom benchmarks, with side-by-side browser replays comparing NanoJev, Jev, and untuned Qwen — driving attention on trending repos.
- 33
disktree is a desktop disk-usage analyzer that scans a disk (your home directory by default) and draws every directory as a nested treemap sized by real disk cost. You navigate with keyboard or mouse, mark folders and files to delete, review the list, then commit — with free space always visible. Built in Rust with GPUI, it integrates with the Omarchy desktop theme.
Why now: It's trending on GitHub's most-starred new repos, gaining over 1,200 stars within roughly a day of its first release — likely driven by its polished treemap UI and Omarchy association.
- 34
AirCard-iOS is a Swift/Rust app for iOS 27+ that customizes Apple Wallet card artwork, lock screen passcode dialer themes, and PosterBoard wallpapers directly on device without jailbreak. It talks to internal system services via a LocalDevVPN loopback tunnel, supports batch card flashing, .passthm theme import/export, .tendies wallpaper installation, and is sideloaded as an IPA via SideStore, AltStore, or TrollStore.
Why now: It's trending as one of the most-starred new repos, likely for showcasing on-device system customization without a jailbreak on iOS 27.
- 35
Laya-CoreML runs the Laya typed-decision model entirely on-device on Apple Silicon via Core ML and the Neural Engine, returning calibrated probabilities for multiple-choice, score, and boolean questions with no token generation. It includes a local Snake game demo, a Python API, and reproducible latency and energy benchmarks (~5 ms per short decision on M3 Max).
Why now: It's trending as a newly popular repo showcasing validated on-device inference with notable energy-efficiency wins over MLX, though the authors note the claimed 10x speedup was not achieved.
- 36
A single-file HTML reader for the open-source Chinese book 《高性价比人生指南》(How to Live Better). It renders all 32 sections and 528 tips as searchable cards showing evidence grade, cost-effectiveness tier, and plain-language explanations. It's fully self-contained—no CDNs or network needed—works offline on phones, includes dark mode, filtering, and daily auto-rebuilds from the upstream book repo.
Why now: It's trending as one of the most-starred new repos, riding interest in the original public-domain book by offering a zero-dependency, mobile-friendly reading version.
- 37
PipePipe is an open-source Android app, forked from NewPipe, for browsing YouTube (and other services) without ads or official app restrictions. It adds SponsorBlock segment skipping, dislike restoration, background music playback, live chat danmaku overlays, keyword/channel filtering, playlist downloads, and gesture-based playback controls.
Why now: It was discussed on Hacker News as a NewPipe hard fork implementing SponsorBlock, which drew attention to its enhanced feature set.
- 38
Reladraw is a text-based diagram language, written in TypeScript, where placement is explicitly stated with relative phrases like 'below' or 'right of', rather than coordinates or auto-layout. It sits between auto-layout tools (Mermaid, Graphviz, D2) and hand-positioning tools (draw.io, Excalidraw), letting you describe a specific arrangement in readable text that re-renders live in the browser.
Why now: It was featured on Hacker News as a Show HN launch, positioning itself as filling a gap between auto-layout languages and manual diagram tools, especially for AI agents editing diagrams.
- 39
A TypeScript trading bot that runs an AI model (Jev) against the Kuru MON-USDC order book on Monad, posting a fresh post-only limit order every ~300 ms block to earn the spread rather than pay it. It streams decisions, quotes, fills, gas costs, and PnL to a dashboard via SSE, with a dry-run mode when no private key is set.
Why now: A brand-new, quickly-starred repo (2.4k stars within a day) demoing AI-driven block-by-block market making on Monad, with a live public dry-run dashboard.
- 40
mini-AGI is an experimental byte-level language model that learns continually from a single stream of data without catastrophic forgetting. It trains from scratch on a single 8GB VRAM GPU, stores weights as files on disk and pages them to the GPU as needed, grows new capacity when short, and prunes unused parts. The README notes it's a toy-level experiment, with weights not yet published.
Why now: It was featured on Hacker News as a 'Show HN' post, likely attracting attention because it demonstrates continual learning on modest consumer hardware — a challenge to the assumption that training requires massive resources.