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.
- 41
This repo holds the source scripts (mostly PowerShell/Packer) that build the virtual machine images used by GitHub-hosted runners for GitHub Actions and Microsoft-hosted agents for Azure Pipelines. It documents available images (Ubuntu, macOS, Windows, arm64 variants), the preinstalled software on each, image release dates, and how to propose changes or build the images yourself.
Why now: Ongoing discussion around image updates and deprecations (e.g., newer Ubuntu/macOS/Xcode images rolling out and older ones being retired), since teams' CI workflows depend on which tools are preinstalled
- 42
Decider is an independent open reproduction of TypeSafe AI's 'System One' (Jev) model class: Qwen3.5-based models that don't generate text. Given a state and typed questions (choice, score, or yes-probability), each model returns calibrated probability distributions for all questions in a single forward pass — no decoding or parsing, with sizes from 2B to 35B MoE.
Why now: A fresh vLLM-based serving release (decider-ai 1.5.0) plus improved 2B/4B checkpoints landed this week, with demo GIFs (Tetris, Pong) and JevBench benchmark numbers driving attention on trending lists.
- 43
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
- 44
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.
- 45
scriptc is an experimental Vercel Labs compiler that turns TypeScript and JavaScript into typed IR, readable C, LLVM IR, assembly, native executables, and WebAssembly. It uses the TypeScript compiler for parsing and type-checking, ships a small native runtime (no Node in output), and falls back to an embedded quickjs-ng engine for npm packages and any-typed code that can't compile statically.
Why now: A Vercel Labs experiment with rapid star growth and active development; attention around compiling TypeScript directly to native code without Node or a JS engine in the output.
- 46
WeChatBridge is a native macOS tool (Swift, sandboxed Share Extension) that adds WeChat's 'forward to other apps' targets so users can send multi-selected chat records—exported by WeChat as a ZIP of text, images, and video—straight to AI agents like Codex, Claude, Doubao, or to Obsidian as Markdown notes, the clipboard, or custom apps. Everything stays local: it reads only WeChat's own exported files, never the WeChat database.
Why now: Recently released a Developer ID signed and notarized 0.1.14 DMG and gaining stars on GitHub trending, riding interest in piping chat context into AI agents.
- 47
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.
- 48
A GitHub repository collecting real AI engineering interview questions from 35 top companies (Anthropic, OpenAI, Google, Meta, NVIDIA, Databricks, Cursor, and more), organized company by company. It also groups questions by topic like LLM internals, inference and GPU performance, RAG, agents, fine-tuning, evaluation, safety, and AI system design, with answers linked where available.
Why now: It recently trended on GitHub as a new, highly-starred repo, reflecting strong demand for AI/LLM engineering roles and interview prep material.
- 49
A curated awesome-list of open-source apps that replace paid SaaS products while running entirely in your own Cloudflare account using Workers, D1, and R2. Entries include Calendly-style booking pages, analytics, feedback boards, inboxes, and more. Each entry is audited: licences are read from licence files and Cloudflare bindings from deploy configs rather than trusting project descriptions.
Why now: It's trending as one of the most starred new repos, riding interest in deploying full apps to Cloudflare's free/cheap tiers instead of paying for SaaS subscriptions or VPS hosting.
- 50
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.
- 51
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.
- 52
TypeLLM is a Python library that adds type-safe, schema-constrained generation to existing autoregressive LLMs served via SGLang, without modifying model weights. It guarantees outputs like strings, integers, booleans, and enums stay within a JSON Schema, with low token overhead, shared-prefix KV caching, dependency graphs between fields, optional per-field thinking mode, image input, and permutation averaging to reduce option-order bias.
Why now: It's trending on GitHub's most-starred new repos, with rapid daily updates, strong JevBench results (228/231 with thinking), and a recent image-input release.
- 53
AnyJev is a Python library that converts any LLM into a calibrated decision model without fine-tuning. It wraps language models to produce typed decisions with real probability distributions (e.g. routing a ticket to billing/technical/sales), fixing order-flip instability and improving calibration. It supports multiple levels, vLLM serving, and fitting a lightweight head with as few as 100-500 labels.
Why now: It recently appeared on GitHub trending as one of the most-starred new repos, with an active vLLM integration making calibrated LLM decision endpoints easy to deploy.
- 54
Plexo is a cross-platform desktop download manager (Windows, macOS, Linux) that speeds up file downloads by splitting them into byte-range chunks and downloading them in parallel over multiple network interfaces at once — such as Wi-Fi, Ethernet, USB-tethered phones, or cellular — aggregating their bandwidth into one faster transfer. It auto-detects network interfaces and dynamically scales worker connections per interface.
Why now: It recently appeared among new, most-starred GitHub repos, likely driven by its demo video and the unusual trick of aggregating bandwidth across physically separate internet connections.
- 55
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.
- 56
Splash is a local LLM inference engine for Apple silicon Macs, written in C++ with Metal kernels and DFlash 2 speculative decoding. It runs coding agents and OpenAI/Anthropic-compatible apps on one machine, with vision, tool calling, a built-in chat page, prefix caching, automatic memory planning, and optional KV cache offload to SSD.
Why now: It's trending on GitHub's most-starred new repos shortly after release, and its README references a fresh DFlash 2 blog post and an LM Studio integration guide, suggesting recent ecosystem buzz.
- 57
Rizzo Flow is an open-source, fully local alternative to TypeSafe's hosted Jev service. It turns unstructured state into typed, probabilistic decisions (yes/no, choices, scores) without generating text tokens, using fine-tuned Spark-X2.5 models on llama.cpp. It exposes a Jev-compatible HTTP API, so existing client code only needs its URL changed to localhost.
Why now: Fresh fine-tuned weights (25 Sept 2026) improved accuracy and probability calibration on the typed-decisions benchmark, and it's trending among newly starred repos as a local, open-weights answer to Jev.
- 58
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.
- 59
A fork of llama.cpp that adds tiered KV-cache memory: completed KV blocks live in host RAM, and a small bounded GPU window is filled with relevant retrieved blocks per step. This lets large-context models like Qwen3.8-27B run at full 256K context on a 16 GiB GPU with near-lossless quality, served via an OpenAI-compatible chat server.
Why now: Recently at the top of trending repos with a new v0.16.0-rc3 prebuilt release, plus benchmark comparisons against KV-cache streaming showing 32-33 tok/s decode on an RTX 5060 Ti.
- 60
GoLive is an open-source Agent Skill plus zero-dependency Node CLI that helps you launch apps built by coding agents. It detects what your project needs, proposes a plan, waits for your approval, applies hosting, database, domain, DNS, email and payments setup using your own provider accounts, then verifies results and can tear everything down. No GoLive account, backend or telemetry.
Why now: Trending on GitHub's most-starred new repos, riding interest in agent-driven deployment: it tackles the gap between an agent building an app and actually shipping it, with an approval-first trust model.
- 61
typesafe-computer-use (jev) is a macOS (Windows experimental) computer-use agent that drives your real mouse and keyboard toward a goal typed in plain English. Instead of asking a frontier model at every step, it OCRs the screen and lets a small TypeSafe classifier pick the next action, reserving a writing model only for free text. README claims ~$0.0002 per decision, with facts computed in code, not by the model.
Why now: It's trending as a newly most-starred repo making a bold cost/speed comparison: the README benchmarks its classifier loop against a frontier model (Claude Opus 5) at 130x–390x cheaper and 3.7x faster per step, while candidly noting reasoning must be rebuilt as deterministic code.
- 62
git-bug is a bug tracker embedded directly in your git repository. Issues are stored as git objects, so you can push and pull them via your normal remotes, work fully offline, and avoid dependence on external services. It offers a CLI, terminal UI, and web UI, plus bridges to sync with other bug trackers and a GraphQL API.
Why now: It was featured on Hacker News, drawing renewed attention to its distributed, offline-first approach to issue tracking.
- 63
virtio-nvgpu is an experimental Rust-based virtio device that forwards NVIDIA kernel driver ioctls between a KVM guest and the host at the driver ABI level. The guest runs unmodified NVIDIA user-mode drivers, targeting headless GPU streaming where a compositor in the VM renders and encodes H.264 on the host's GPU. Benchmarks show guest performance within ~2% of bare metal for game-sized frames, with several guests sharing one card.
Why now: It was featured on Hacker News, and it's remarkable because it claims near-native (within 2%) NVIDIA GPU performance in VMs without GPU passthrough — long a pain point for cloud gaming and virtualization.
- 64
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.
- 65
Shapeshift is a TypeScript UI component where a single text box dynamically morphs into the right interface as you type — an event card, checklist, timer, color picker, bill splitter, converter, poll, and more. Intent classification is done by the TypeSafe Jev model (or a built-in offline keyword classifier), while dates, amounts, and math are handled by deterministic code.
Why now: It's newly created and rapidly starred on GitHub's trending list, with a live demo and a 60-second video showing the morphing input in action.
- 66
Von is an open-source, non-autoregressive 'System One' decision model for fast classification, routing, and guardrail-style tasks. Instead of generating text token-by-token, it scores discrete/probabilistic options in a single forward pass, delivering sub-25ms decisions with calibrated uncertainty. Version 1.2 adds order-invariant option scoring, fixing answer flips under option reordering, plus Intel GPU acceleration via OpenVINO.
Why now: A fresh 1.2 release fixing an order-sensitivity defect, demonstrated with a Doom gameplay demo, is drawing attention on the most-starred repos list.
- 67
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.
- 68
AirCard is a native Windows tool (Rust, ~7.5 MB single executable) that lets iOS 18+ users customize Apple Wallet card faces and apply lockscreen passcode keypad themes without jailbreaking. It detects cards in real time, applies custom artwork over USB or WiFi, and restores the original state afterward using a sync-based exploit.
Why now: It is trending among newly starred repos, likely due to offering no-jailbreak iOS customization on Windows with a lightweight native client.
- 69
A set of Python-based agent 'skills' that offload small, frequent decisions from expensive frontier LLMs to Jev, TypeSafe's decision-only model. It handles model routing, memory filtering, prompt-injection screening, transcript compaction, skill selection, message triage, mailbox sorting, and computer/browser use. Skills ship as plain SKILL.md files, so they work with Hermes agents, Claude Code, Codex, or anything that reads skill files.
Why now: Recently created and climbing the most-starred new repos list, with freshly published evaluation scorecards showing strong injection-detection and recall results.
- 70
text-humanizer is a free, open-source Python tool that rewrites AI-generated text to sound more natural and evade AI detectors. It runs a multilingual pipeline: DeepSeek rewrites the text via a Chinese intermediate representation, optional Google Translate and DeepL steps add cross-lingual variation, and a final DeepSeek pass reconstructs fluent text in the original language across 8 supported languages.
Why now: It recently appeared among the most-starred new repositories, gaining attention for openly advertising AI-detector bypass (ZeroGPT and similar), a topic that draws both interest and controversy.
- 71
Awesome Jev is a community-maintained, MIT-licensed catalog listing 916 verified open-source projects built with Jev, TypeSafe AI's model for typed decisions in software. Entries are organized into 11 categories (agent tooling, applications, benchmarks, etc.) across 27 languages, with each listing pointing to concrete evidence of how Jev is actually used.
Why now: It's trending as one of the most-starred new repos, actively updated (20 projects added September 25) amid growing ecosystem interest around Jev and Jev-like models.
- 72
Dream-RSI is a research project (paper from Google, DeepMind, UMD, UVA) on recursive self-improvement of AI agents. Its key idea is to use accumulated discovery history as a replay simulator so the agent can improve its own exploration strategy, evaluated across 8 tasks in algorithm engineering, math optimization, and GPU kernel engineering, with results like 1.22x faster downstream runtime.
Why now: It's a freshly released preprint (Sep 2026) from Google/DeepMind researchers that quickly gained ~1200 stars, and it claims progress on recursive self-improvement — a hot and controversial topic in AI research.
- 73
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.
- 74
Astra Flash Orchestrator is a Python-based Codex workflow that splits AI coding labor by cost: Astra handles planning, architecture, and final review, while the cheaper DeepSeek V4.1 Flash does the bulk implementation, testing, and debugging. It packages work into coherent task bundles, collects evidence, and includes reversible, dry-run installation, with OpenCode support.
Why now: It's trending on GitHub's most-starred new repos, boosted by benchmark claims of ~98% lower cost per 1K lines versus an all-Astra workflow.
- 75
Qwen-Image-2.1 is Alibaba Qwen's newly open-sourced text-to-image generation and editing model. Its 7B-parameter DiT-based visual component generates regular or transparent (RGBA) images, supports editing with up to 10 reference images, local edit regions via masks or annotations, identity preservation, and improved typography and textures, all while staying computationally efficient.
Why now: Freshly released (Sept 20, 2026) with open weights and Day-0 support in Diffusers, ComfyUI, vLLM-Omni, SGLang, and LightX2V; it's trending as a top-starred new repo.
- 76
DeepOpen is an open-source, non-autoregressive 'System 1' decision engine for structured decision tasks like classification, routing, and scoring. Instead of generating text token by token, it produces outputs in a single forward pass, supporting 100+ languages with millisecond-level latency, built on three specialized checkpoints with an automatic router, trained via RLCD reinforcement learning.
Why now: It recently appeared on trending 'most starred new repos' lists and claims strong leaderboard results on the Banking77 and CLINC150 intent-classification benchmarks, with reproducible one-click pipelines.
- 77
A complete Lean 4 formalization of Michael Spivak's classic textbook Calculus, covering every definition, theorem, and problem from all 30 chapters and 9 appendices of both the 3rd and 4th editions. It uses Spivak's own definitions, proves results like the transcendence of e and π (some parts missing from Mathlib are supplied), and documents about sixty statements in the book that are false as printed.
Why now: It was shared on Hacker News as a Show HN post, notable for its completeness and for formally flagging errors in Spivak's printed text — including sixteen answer-section mistakes.
- 78
AgentRun is a JavaScript/TypeScript DSL for defining AI agent workflows as explicit, repeatable steps. Instead of letting an agent improvise, you declare nodes for tool calls, focused decisions (via a 'Jev' judge), and agent invocations, with schemas, deadlines, thresholds, and escalation paths. It ships as an npm interpreter with a TypeScript builder, Zod contracts, parallel mapping, loops, and sub-workflows.
Why now: It was just launched on Hacker News as a Show HN, positioning itself as a structured alternative to free-form agent loops.
- 79
A demo repo showcasing Claude Opus 5.5's code-generation ability: three playable 3D web games (a pelican biking game, an FPS map recreation, and a drift racing game) each generated from a single one-sentence prompt in one session, with zero human edits to the code. Games are built with Three.js and esbuild into single-file HTML pages deployed on Cloudflare Pages.
Why now: It's trending as a viral one-shot demo of Claude Opus 5.5's capabilities, with playable deployed games serving as proof.
- 80
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.