From Product Hunt
Developer tools dominated today. A visual feedback layer for AI agents, Stripe's full-stack CLI, and a couple of solid macOS utilities.
Agentation — Visual feedback tool for AI agents
Give Claude Code, Codex, and other AI agents visual context. Click UI elements, annotate screenshots, and paste structured output back to your agent. MCP integration built in. Free for individuals.
Stripe Projects — Production-ready dev stack from CLI
Provision hosting, databases, auth, AI, and analytics from the command line. Credentials sync automatically. Everything billed through a single Stripe account.
Noctiluca — Remote desktop for macOS on custom QUIC
Built on Sirius, an open-source protocol using custom QUIC. Hardware-accelerated H.265/H.264 encoding, multi-display support, and SSH key authentication. Free cross-platform clients.
DashPane — Lightning-fast window switcher for macOS
Fuzzy search across individual windows, not just apps. Finally a switcher that understands you have 12 browser windows open. $4.99 one-time.
1DevTool — Multi-project IDE with persistent terminals
Persistent tmux terminals, built-in browser, API client, database viewer, and AI agents all sharing context. One workspace for everything.
Stakpak Autopilot — Open-source self-healing infra agent
A Rust binary that watches your apps, auto-fixes issues, and escalates to humans when it can't. Cedar policy engine for guardrails. Runs Docker-sandboxed.
From Reddit
Big one today. An accidental data leak revealed Anthropic's next model, GLM-5.1 is closing the gap with frontier models, and TurboQuant is making waves in the local LLM scene.
Fortune exclusively reported that Anthropic accidentally left details of an unreleased model called "Claude Mythos" (codename "Capybara") in a publicly accessible CMS data store. Around 3,000 unpublished assets were exposed.
An Anthropic spokesperson confirmed it represents "a step change" in performance and is "the most capable we've built to date," with "unprecedented cybersecurity risks." Benchmarks show dramatically higher scores on coding, reasoning, and cybersecurity tasks compared to Opus 4.6. Currently in early access testing.
GLM-5.1 released — 94.6% of Opus 4.6 on coding benchmarks
Zhipu AI's new flagship: 744B total parameters, 40B active (MoE). 200K context window, 128K output. Trained entirely on Huawei Ascend chips. Weights dropping April 6-7. Coding plan starts at $3/month.
Judge blocks Pentagon's Anthropic supply chain risk label
Update: Federal Judge Rita Lin issued a 43-page ruling indefinitely blocking the designation. Called it "classic illegal First Amendment retaliation" and rejected the "Orwellian notion that an American company may be branded a potential adversary for expressing disagreement with the government."
TurboQuant in llama.cpp: +22.8% decode speed at 32K context
Developer implemented TurboQuant for KV cache in llama.cpp. On M5 Max at 32K context, dequant was eating 40% of decode time. After trying 14 different approaches, Flash Attention's tiled access pattern lets you skip 90% of dequant work.
AI bots officially overtaking humans on the internet
Human Security's "State of AI Traffic" report: automated traffic grew 8x faster than human activity, with AI-driven traffic surging 187% year-over-year. Cloudflare's CEO predicts bots will exceed human traffic by 2027.
Quick hits from the local LLM scene.
TurboQuant enables Qwen 3.5-9B on MacBook Air M4 16GB
20K context on hardware that couldn't run it before. Patched llama.cpp makes it possible.
Unsloth Studio ships 50+ features in one week
Pre-compiled llama.cpp binaries, auto-detection of local models, 20-30% faster inference.
From Reddit
Interactive world models are getting real, LTX keeps improving, and someone built proper film emulation for ComfyUI.
Matrix-Game 3.0 — Real-time interactive world models
Skywork AI's new release runs at 720p @ 40FPS with the 5B model. MIT license. Minute-long memory consistency using Unreal, AAA, and real-world training data. Scales up to 28B MoE.
LTX 2.3 running real-time on a single RTX 4090
Achieved using Scope, an open-source real-time AI pipeline tool. Real-time video generation on consumer hardware is getting closer to practical.
LTX 2.3 audio fix: swap scheduler for clean output
Replacing LTXV's built-in scheduler with BasicScheduler fixes the metallic hiss issue.
ComfyUI-Darkroom — Real film emulation from Capture One data
11 nodes with 161 film stocks parsed from 586 real Capture One XML curve files. Color and B&W separation.
From Reddit
A massive self-hosting guide went viral, some great new homelab tools, and a macOS notch app that actually looks useful.
Free 750-page self-hosting production guide
Comprehensive guide from a decade-long self-hoster covering everything from basics to handling production traffic. Hand-written, no AI slop. The kind of resource you bookmark permanently.
Homelable — Interactive homelab infrastructure visualizer
Live network diagrams with real-time status monitoring. Includes MCP integration for AI clients. Docker one-line install.
NOMAD — Self-hosted trip planner
Real-time collaboration, interactive maps, budgets, and packing lists. Docker deployment. Replaces the mess of Google Docs and WhatsApp group chats for trip planning.
Droppy — Turn the macOS notch into a productivity hub
Native app with file shelf, floating basket, clipboard manager, media player, AirPods HUD, volume/brightness controls, and 16+ extensions. Lifetime purchase.
Rangarr — Security-hardened Huntarr replacement
Ground-up rewrite after Huntarr's security meltdown. No telemetry, three auditable source files.
Subscription Day — Track paid subscriptions
Completely rewritten in Swift. iOS and iPad with sync. Lifetime purchase.
Netfluss — Open-source menubar network utility
DNS switcher with Unifi, OpenWRT, and Fritzbox router support. Free and lightweight.