From Product Hunt

Agent tooling dominated today's launches. Open-source runtimes, self-hosted agents, browser bridges, and shared context layers across AI tools.

OpenHuman — Open Source AI Agent Harness Local-first AI agent with persistent memory across sessions. Encrypted vault, one-click setup, connects to Gmail/Slack/Telegram/Notion/GitHub. GNU licensed. 8K+ GitHub stars in its first week.
▲ 392 · tinyhumans.ai
TrustClaw — Self-Hosted AI Agent on Vercel MIT-licensed personal AI agent that connects to 1,000+ apps over OAuth via Composio. Runs scheduled jobs, talks to you on Telegram, deploys with one click. Your tokens never leave your infra.
▲ 134 · github.com
Cline SDK — Open-Source Agent Runtime TypeScript agent runtime with plugin architecture, native subagents, MCP support, and checkpointing. The same runtime Cline uses internally, now available as an open package. Provider-agnostic.
▲ 105 · cline.bot
Kimi WebBridge — Browser Control for AI Agents Browser extension that lets terminal agents operate your Chrome/Edge session. Opens pages, fills forms, extracts data. Works with Claude Code, Codex, Hermes, and OpenClaw. Local-first, no tokens sent to third parties.
▲ 101 · kimi.com
Relay — Shared Context Across AI Tools Keeps a living project brief that follows you across ChatGPT, Claude, Gemini, and Grok. One-click context injection into new conversations. MCP integration syncs with Cursor, Claude Code, and Codex.
▲ 102 · onrelay.app
Picsart MCP — 140+ AI Models via Single MCP Unified MCP server for image, video, and audio generation. One connection to access 140+ models. Describe what you need in plain English and the server picks the right model. Works with Claude and ChatGPT.
▲ 88 · picsart.com
Cats Lock — Lock Your Mac Keyboard for Cats Tiny macOS utility built for the cat-on-keyboard moment. Cmd+L toggle, stealth mode for meetings, customizable warning sounds. Sandboxed, no analytics. $2.99 one-time purchase.
▲ 89 · catslock.app
From Reddit

OpenAI reshuffles again, considers suing Apple, and launches personal finance. Meanwhile, Mythos continues to terrify the security world and a new inference technique delivers nearly 8x speedups.

OpenAI Considers Legal Action Against Apple Over ChatGPT Integration OpenAI has hired outside counsel, complaining that Apple buried the ChatGPT integration in Siri, made features hard to find, and that revenue is "nowhere close to projections." Legal move likely after the Musk trial concludes.
ChatGPT Personal Finance Launches for Pro Users Connect bank accounts via Plaid to 12,000+ institutions including Chase, Fidelity, and Robinhood. Get a dashboard of portfolio performance, spending patterns, subscriptions, and upcoming payments. US Pro users only for now.
AI Breaks Every Benchmark for Autonomous Cyber Capability New benchmark data shows Mythos Preview massively outperforms GPT-5.5 on n-day exploit tasks. Mythos solved 157 test instances vs GPT-5.5's 120 in two hours. Open-weight models scored zero. On the hardest tasks, GPT-5.5 scored 71.4% vs Mythos's 68.6% — within margin of error.
Orthrus: 7.8x Faster Inference with Provably Identical Output New technique generates up to 7.8 tokens per forward pass on Qwen3-8B by training a lightweight multi-token prediction head on top of a frozen backbone. Zero quality degradation — mathematically identical output distribution. Models available on HuggingFace.
Intern-S2-Preview: 35B Scientific Multimodal Model Shanghai AI Lab released a 35B scientific foundation model that explores "task scaling" — increasing difficulty and diversity of scientific tasks. Achieves performance comparable to frontier models on research-oriented benchmarks.
Fully Offline Suitcase Robot Running Gemma 4 on Jetson Orin NX Impressive build: 30+ sensors, SenseVoice STT, Piper TTS, PixiJS face display, ~200ms cached TTFT at 14-15 tok/s. No WiFi, no cloud. Vision and OCR native to Gemma 4.
From Reddit

Microsoft leaked (then deleted) a new image model, and the community built multi-GPU inference over regular Ethernet.

Microsoft Accidentally Uploads "Lens" Image Model to HuggingFace Microsoft briefly published Lens and Lens-Turbo — a 3.8B parameter text-to-image model with 1440x1440 output. The repos were quickly deleted but not before the community grabbed details. Smaller than FLUX, potentially runnable on 12GB cards.
Pixal3D — High-Fidelity 3D Assets from a Single Image TencentARC's new model generates detailed 3D meshes with PBR textures from one photo. Uses back-projection for direct pixel-to-3D correspondence instead of loose attention injection. Runs locally.
ComfyUI-Mesh: Split FLUX Across GPUs Over Ethernet LAN Custom NVENC bridge that distributes FLUX model layers across machines over regular Ethernet. 5090 + laptop 4090 = 4.4s per image. Works over WiFi 6 and even Tailscale. No NVLink needed.
From Reddit

Bitwarden's free tier may be on borrowed time. Plus a few macOS utilities and a self-hosted bookmark manager worth checking out.

Raybeam — Better Screen Sharing on macOS Native Mac app for sharing specific screen regions instead of entire monitors. Built for ultra-wide setups where sharing the full display is impractical. Supports multiple regions, cursor tracking, and picture-in-picture preview.
Faved — Self-Hosted Bookmark Manager with Nested Tags Open-source bookmark manager designed for large collections. Organize with hierarchical nested tags (e.g., Programming Languages > Backend > Go). Build full taxonomies instead of flat tag lists. Self-hosted, Docker-ready.
Sidebar Calendar — Always-Visible Mac Calendar Persistent sidebar that shows your schedule without opening a separate app. 70 updates in one year of active development.
Kavita v0.9.0.2 — Critical Security Patch All prior versions affected. Self-hosters should update immediately. CVE details to be shared after users have time to patch.