From Product Hunt
Claude gets a phone remote, MCP keeps expanding its ecosystem, and Unsloth makes local LLM training accessible to everyone.
Claude Dispatch — Text Claude from your phone, it runs on your desktop
New Cowork research preview. Message Claude from your phone — it accesses your files, browser, and local tools on your desktop. Sandboxed and local, you approve before it acts. One persistent conversation. Max plans first, Pro soon.
Unsloth Studio — Open-source web UI to train and run LLMs locally
No-code UI for training, running, and exporting 500+ models locally. Transforms PDFs and CSVs into datasets, fine-tunes 2x faster with 70% less VRAM. macOS, Windows, Linux. 100% offline.
OpenObserve — Open-source Datadog alternative in Rust
Observability platform for logs, metrics, and traces with 140x lower storage costs than Elasticsearch. Stateless architecture, S3-compatible storage. Self-hostable, start in 2 minutes.
Permit.io MCP Gateway — Zero-trust security for any MCP server
Drop-in proxy adding OAuth auth, fine-grained authorization, consent screens, and audit logging to any MCP server. One URL change, no code modifications. Tracks full human-to-agent delegation chain.
Fantastical MCP for Claude — Manage your calendar inside Claude
MCP connector brings Fantastical's natural language scheduling into Claude. Create, reschedule, and delete events without switching apps. Handles time zones, recurring events, and vague requests natively.
Lore — Local AI second brain, 100% private and open-source
System tray app for capturing and recalling ideas using Ollama + LanceDB for offline RAG. Ask questions in plain language, get answers from your own history. No API keys, no cloud, zero tracking.
Forvibe for macOS — Everything between your build and the App Store
Native Mac app streamlining App Store launches. Localization across 40+ languages, screenshot generation, IAP pricing, landing pages, legal docs, and customer support inbox. All from your dock.
From Reddit
GTC 2026 dominates the news cycle with Nvidia's agent platform and Nemotron models. Harmonic's autonomous mathematician goes free, and OpenAI turns model compression into a talent hunt.
Free agent that interprets math problems in plain English, converts them to formal Lean proofs, and works continuously for up to 24 hours. Ranks #1 on ProofBench, outperforming closest competitor by 15%.
Co-founded by Robinhood CEO Vlad Tenev. This is the tool behind the recent Erdős problem news that Terence Tao attempted to solve using ChatGPT.
Nvidia launches open-source agent platform at GTC 2026
Three components: NemoClaw secure runtime with policy-based guardrails, AI-Q research agent blueprint topping DeepResearch Bench, and Nemotron open models. Partners include Adobe, Atlassian, SAP, Salesforce, and others.
OpenAI Parameter Golf — train the best model under 16MB
Public challenge: build the strongest LM where weights + training code fit under 16MB, trained in 10 min on 8 H100s. $1M in RunPod compute credits. Top performers get job interviews. Runs March 18 to April 30.
Nemotron 3 Nano 4B — compact hybrid model for efficient local AI
Nvidia's 4B parameter hybrid combining transformer and Mamba-2 architectures. Optimized for edge deployment and consumer hardware. Part of the GTC 2026 Nemotron family announcements.
Britannica and Merriam-Webster sue OpenAI for "massive" copyright infringement
Publishers allege ChatGPT cannibalizes their traffic and ad revenue. Filed in SDNY, claiming OpenAI built its $730B company on the back of their researched content.
Quick hits from AI this week.
Pentagon plans secure environments for AI training on classified data
MIT Tech Review: AI companies would train military-specific model versions on classified data. Follows existing Claude use for analyzing targets in classified settings.
Flash-MoE runs Qwen3.5 397B at 5.7 t/s on 48GB M3 MacBook
Built on Karpathy's autoresearch and Apple's "LLM in a Flash" paper. Math suggests 18 t/s is achievable on the same hardware.
From Reddit
Local video generation tools keep maturing — two new open-source editors aim to bridge ComfyUI workflows with practical video production.
vlo — ComfyUI-powered local open-source video editor (alpha)
Browser-based video editor that bridges directly with ComfyUI. Any ComfyUI workflow should be compatible. Combines traditional timeline editing with AI generation pipelines in one interface.
Synesthesia — 100% local AI music video generator
Feed it a vocal stem, full performance, and lyrics. Local LLM (Qwen3.5-9B) creates a plotline, then LTX Video Desktop generates the clips. Auto-cuts between vocal performance and story scenes.
FlashMotion — 50x faster controllable video generation
Few-step gen on Wan2.2-TI2V with multi-object box/mask guidance and camera motion. Weights on HuggingFace.
From Reddit
Mac utilities, a creative reference tool, and a self-hosted MinIO replacement.
Refbox — floating references that stay above your apps
Mac app for keeping images, videos, GIFs, and Markdown notes visible while you work. Masonry mode, gallery mode, frame-by-frame video playback. Built for designers, artists, and creatives.
Chronoid — automatic time tracking with MCP server for Claude
Runs in the background logging every app, website, and document. New MCP server lets Claude query your time data via conversation. One-time purchase $20, fully local.
Firemark — anti-AI watermarking CLI in Rust
Open-source CLI for watermarking images and PDFs with anti-AI removal defenses. Cryptographic guilloché/moiré patterns, non-deterministic perturbation, and adversarial metadata injection.
OpenTeleprompter — voice-activated, invisible during screen share
Free, open-source Mac teleprompter. Voice-activated scrolling that automatically hides when screen sharing on Zoom, Meet, or Teams. Fully local, no account needed.
RustFS — open-source MinIO replacement with binary migration
S3-compatible storage replacing the archived MinIO. Migrate via simple binary swap, no data re-indexing needed.