217 projects tracked · Updated Apr 4
Anthropic blocks Claude Code subscriptions from accessing OpenClaw (914 HN pts, 708 comments, 0.77:1 ratio) — OpenClaw has 347K GitHub stars, surpassing React's 243K to become the most-starred active software repository in history, and now Anthropic is terminating its own product's access to it. The 0.77 comment-to-point ratio is the week's highest controversy signal; the community is not debating facts, it is processing what a closed platform tax means when the ecosystem has already escaped into the wild.
CVE-2026-33579 — OpenClaw privilege escalation vulnerability surfaces on NVD (476 HN pts, 221 comments) the same week Anthropic bans OpenClaw access. The temporal proximity creates interpretive ambiguity that will define the narrative for weeks: responsible disclosure or strategic timing? Security incidents that coincide with platform disputes are no longer treated as coincidences by the practitioner community.
Artemis II crew photograph Earth from 270,000 miles — day 4 of mission generates 919 HN pts and 318 comments, more engagement than the launch. The community is not reacting; it is archiving. Each mission update becomes a collective memory document for a generation watching the first crewed deep-space mission in 53 years unfold in real time.
Claude Code finds a 23-year-old Linux kernel vulnerability autonomously (214 HN pts, 130 comments; 67 Lobsters pts, 19 comments) — cross-platform confirmation is the reliability signal: when a story clears both HN and Lobsters with separate organic uptake, it has passed a high practitioner-credibility bar. Autonomous AI as kernel-level vulnerability scanner is now demonstrated fact, not speculation. The security research community's skepticism window is closing.
oh-my-codex reaches +1,803 stars/day (15.3K total, TypeScript) with hooks, agent teams, and HUDs layered on OpenAI Codex — the enhancement-shell pattern is now a standalone product category. Every major AI coding tool spawns a community project that does what the parent refuses to: oh-my-codex → codex; superpowers → claude-code; the pattern predicts openclaw enhancement layers will follow identical trajectories. The open-source immune response to closed platforms is a shell, not a fork.
AI agent skill that researches any topic across Reddit, X, YouTube, HN, Polymarket, and the web, then synthesizes a grounded summary. Watchlist mode for recurring research. #1 GitHub trending at 10,211 stars (+2,685/day). First breakout 'single-purpose agent skill' — not a framework.
Meta's self-referential self-improving agents — a meta agent modifies both the task agent AND its own improvement procedure. Cross-domain transfer (robotics strategies improving math grading). 121 HN pts. CC non-commercial license.
Autonomous CLI agent for deep financial research — task planning, self-reflection, real-time market data. TypeScript/Ink/LangChain. Self-validating with LLM-as-judge evaluation suite. 18,972 stars. Second vertical agent breakout after pentagi (security).
Provision and manage production-ready dev stacks from the CLI — hosting, databases, identity, analytics, AI. Integrates Vercel, Railway, Supabase, Neon, PlanetScale, Turso. Credential sync across local/CI/team. 102 HN pts on launch day.
SQLite VFS in Rust serving sub-250ms cold JOIN queries from S3. Page-level zstd compression and AES-256 encryption. Ships as Rust lib, SQLite extension, and Python/Node/Go packages. 101 HN pts (Show HN).
4B parameter OCR model — SOTA 85.9% on olmOCR benchmark. Complex tables, forms, handwriting with full layout. Outputs HTML/Markdown/JSON. Smaller and more accurate than Chandra 1 (9B). 6,130 stars (+557/day). Apache 2.0.
WiFi DensePose — commodity WiFi signals into real-time human pose estimation, vital sign monitoring, and presence detection without video. Rust core, ESP32 mesh ($1/node), Docker deployment. 43,136 stars (+1,002/day). Camera-free privacy-preserving sensing.
Chat platform built on the AT Protocol for communities big and small. AT Protocol expanding beyond microblogging into chat. 98 HN pts.
AI memory engine — persistent memory and knowledge management for agents. Accelerated from +344/day to +810/day (18,863 stars). Memory/synthesis layers becoming first-class demand centers.
Automated content creation and monetization pipeline — AI-generated short videos with multi-platform publishing. 25,139 stars (+3,006/day). Content automation as a distinct agent category.
Markdown-to-email publishing CLI — write emails in Markdown, send as formatted HTML. 283 HN pts. Classic friction-removal for solo devs and agent pipelines.
Faster brew-compatible macOS package manager — addresses daily friction of Homebrew latency. 203 HN pts, 124 comments with deep technical engagement.
Apple Silicon local inference runtime — native CLI for running LLMs on M-series chips. 201 HN pts. Part of the local inference convergence wave.
3D architectural creative tooling — browser-based spatial editor. 6,127 stars (+1,449/day). Non-AI creative tool capturing attention alongside agent frameworks.
Major media player rewrite — 88% smaller bundle size. Anti-bloat simplification earns 380 HN pts. Proves non-AI creative tooling can still capture attention.
ByteDance's SuperAgent harness — researches, codes, and creates using sandboxes, memories, tools, skills, and subagents. Accelerated to 48,267 stars (+2,394/day) by Mar 27, up from 36.6K debut on Mar 23. Fastest-growing multi-agent framework this week.
Fully autonomous AI pentesting system in Go — performs complex penetration testing tasks without human oversight. 12,400 stars (+1,069/day). Go implementation signals production readiness. Fastest-growing security-focused agent.
Claude Code skills that turn codebases into interactive knowledge graphs — shift from 'AI writes code' to 'AI understands code.' Multi-platform (works with Codex). 4,700 stars (+489/day).
48 AI agents mirroring real game studio hierarchy — 36 workflow skills with full coordination system. Most complex multi-agent coordination system documented on GitHub. 2,500 stars (+328/day).
Tiny CLI to see and kill whatever's running on localhost — utility tool embodying the 'small, focused CLI tool' philosophy. 194 HN pts on Show HN.
GitHub's official spec-driven development toolkit — when the platform owner releases tooling for a methodology, adoption becomes institutional. 80,900 stars (+340/day). Validates the 'sufficiently detailed spec is code' thesis.
Design intelligence skill for Claude Code — AI-driven UI/UX design decisions and component generation. 48,300 stars (+213/day). Part of the Claude Code skills explosion.
Bram Cohen's (BitTorrent creator) next-generation version control system — 500 HN pts, 279 comments. Most significant VCS discussion since Git's adoption. Proposes fundamental rethinking of version control.
Autonomous development loop for Claude Code — self-referential execution with exit detection. 8,066 stars. Part of the Claude Code ecosystem expansion beyond plugins into workflow automation.
Agent memory with SOTA local GPU retrieval — GPU-accelerated vector store for agent context. Part of the Mar 22 agent infrastructure gold rush (4 independent memory/context tools in one day).
Compress codebase context for LLMs by 78% using Rust — reduces context window consumption without lossy compression. Complementary to memory tools: compress first, then store/retrieve.
Agent-vs-agent security testing — 3 Claude agents (Red vs Blue) hack and patch code autonomously. Novel paradigm: using AI agents to red-team code produced by other AI agents before deployment.
Agentic Linux kernel code review system — AI-powered review of kernel patches. First domain-specific agent for OS kernel development.
Pure C/Metal inference engine running Qwen3.5-397B (397 billion parameters) on a MacBook Pro with 48GB RAM at 4.4+ tokens/sec. No Python, no frameworks — just C and hand-tuned Metal shaders. The entire 209GB model streams from SSD.
Local-first observability for LLM agents — 'Datadog for agents.' Logging, tracing, metrics, and alerting for agent workloads. If agents are services, they need observability infrastructure.
Professional video editing in the browser — Rust/WebGPU/WASM with near-native performance. Multi-track timeline, keyframe animation, real-time effects. Local-first via File System Access API. Zero installation.
Free, open-source local AWS emulator — emerged after LocalStack's community edition sunset (auth tokens, dropped CI support, frozen security updates). No account, no feature gates, no telemetry. Just `docker compose up`. 249 HN pts.
Multi-agent LLM financial trading framework — agents analyze market data, execute trades, manage risk. 41,343 stars (+1,760/day), with Chinese adaptation TradingAgents-CN at 21K. Domain-specific multi-agent validated.
Minimum viable terminal emulator built on libghostty C API — single C file using Raylib for rendering. Demonstrates Ghostty's VT parser as a standalone library: SIMD-optimized parsing, leading Unicode support, fuzzed codebase. 326 HN pts, 67 comments.
Terminal-first 2D sandbox survival game written in Rust — Show HN with strong engagement. Games-in-terminal as creative expression.
Self-hosted AI accounting app — LLM-powered receipt and invoice analysis for solo developers and freelancers. 2,135 stars at 137/day. AI + self-hosted + financial requires high trust.
Pure-Rust high-performance graph database — embeddable or standalone. Supports LPG + RDF, all major query languages (GQL, Cypher, Gremlin, GraphQL, SPARQL). Zero external dependencies. Fastest on LDBC Social Network Benchmark. Multi-language bindings.
Open-source Fujifilm X RAW STUDIO webapp clone — apply Fujifilm film simulation recipes in-browser. 160 HN pts with enthusiastic photographer community engagement.
Together.ai's state space model — outperforms Transformers by 4% on language modeling, runs 7x faster. Half the state size of Mamba-2. Apache 2.0. Published at ICLR 2026. 292 HN pts. Inference efficiency as the primary design goal.
Meta-prompting, context engineering, and spec-driven development system for Claude Code — competing with superpowers for the 'canonical agentic methodology' position. 39,200 stars, sustained at +301/day.
Agent harness performance optimization system — 13 expert agents, 40+ skills, 32 commands. By Anthropic hackathon winner. Crossed 100K milestone (~108,000 stars). Second Claude Code ecosystem project (after superpowers) to hit 100K.
MoonshotAI/Kimi's drop-in replacement for standard residual connections in Transformers — depth-wise attention over previous layers. Achieves same performance with 1.25x less compute. 236 HN pts. Published at ICLR 2026.
Bluesky / AT Protocol client written entirely in Fortran — Show HN with 137 pts. Fortran resurgence meets decentralized social protocols.
Generate any real-world location in Minecraft from OpenStreetMap data — Rust-based world generation. 12,900 stars (+582/day), sustained multi-day trending. Creative/gaming tool demonstrating Rust for data-intensive generation.
Y Combinator CEO's exact Claude Code setup — 15 opinionated tools that serve as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA. 39,300+ stars. TechCrunch coverage ('Why Garry Tan's Claude Code setup has gotten so much love, and hate'). v0.12.4.0 released.
LangChain's open-source framework for building internal coding agents — triggers from Slack/Linear/GitHub, runs in isolated cloud sandboxes, auto-creates PRs. Sustained growth past initial spike.
Three open-source TTS models — smallest under 25MB, CPU-only, ONNX-based. Sustained to 554 HN pts (Day 3+). Sub-25MB footprint enables bundling with CLI agents for offline voice. Viable piper-tts alternative.
Autoresearch for SAT solvers — AI-driven research agent that autonomously improves SAT solver performance. Karpathy's overnight-experiment paradigm applied beyond ML.
Simple CLI for orchestrating Claude Code — lightweight task runner for multi-step Claude Code workflows. Surged to 283 pts, 87 comments on HN. The Claude Code platform toolchain solidifies.
GPU-accelerated infinite-canvas terminal in Rust — spatial 2D canvas replaces linear scrollback. Pin, zoom, and navigate multiple agent outputs spatially. 76 HN pts.
PDF parser for AI-ready data — automates PDF accessibility and extraction for LLM pipelines. Trending #1 on GitHub at 5,497 stars (+1,394/day).
Random-access JSON alternative — new data format addressing JSON's limitations for large datasets. 133 HN pts, 92 comments with deep technical discussion.
Showing 7 most recent days · 217 total projects across 4 categories