From Product Hunt

AI agent hosting, Google's Opal 2.0, and a couple of clever tools built specifically for Claude Code users.

KiloClaw — Hosted OpenClaw. No Mac mini required. Fully managed, hosted version of OpenClaw—the most popular open-source AI agent. One-click deploy, 500+ AI models, 50+ chat platforms. Handles infrastructure, security, updates, and monitoring.
▲ 556 · kilo.ai
Opal 2.0 — Smart agent, memory, routing and interactive chat Google Labs' no-code AI workflow builder gets a major upgrade. New agent step analyzes your goal and auto-calls tools like Veo for video or web search. Memory persists context across sessions.
▲ 262 · opal.google
Mercury 2 — Fastest reasoning LLM, 1,000+ tokens/sec First reasoning diffusion LLM. Generates tokens in parallel instead of sequentially. 128K context, tunable reasoning, native tool use. $0.25/1M input, $0.75/1M output tokens. OpenAI API compatible.
▲ 133 · inceptionlabs.ai
VibePad — Control AI coding assistants with a gamepad Free macOS menu bar app that maps your gamepad to keyboard shortcuts. Built for Claude Code and Codex. X=approve, O=reject, L2=hold-to-talk dictation. Open source, native Swift.
▲ 116 · vibepad.now
PeonPing — Stop babysitting Claude Code Sound notifications for any AI agent. Plays game character sounds when your coding agent finishes, errors out, or needs approval. 100+ sound packs (Warcraft, StarCraft, GLaDOS, TF2). Desktop notifications and animated ork tamagotchi.
▲ 110 · peonping.com
Thinglo — Save anything from any app, AI-organized iOS save-everything app. Share from any app, AI auto-titles and categorizes. "For You" feed resurfaces forgotten saves. Document scanner with OCR. 100% private—Face ID, on-device, no tracking.
▲ 94 · apps.apple.com
From Reddit

Breakthrough hallucination research, Sonnet 4.6's identity crisis, Karpathy on the coding agent inflection point, and Perplexity's new multi-model computer system.

Sonnet 4.6 Says "I am DeepSeek-V3" When Asked in Chinese Multiple users report that when asked "what model are you" in Chinese, Claude Sonnet 4.6 responds saying it's DeepSeek-V3 developed by DeepSeek. Fuels the distillation accusations Anthropic recently raised against Chinese labs.
Karpathy Says Coding Agents Crossed a Reliability Threshold Andrej Karpathy says coding agents crossed a reliability threshold in December and can now handle long, multi-step tasks autonomously. Describes this as a major shift from writing code to orchestrating AI agents.
GPT 5.3 Codex Tops Agentic Coding, Surpasses Opus 4.6 Codex 5.3 surpasses Opus 4.6 to claim top spot in agentic coding benchmarks. "Blazingly fast" but the xHigh version can be very expensive. Overall global average score still lags behind Opus 4.6, the current leader.
Perplexity Computer — Multi-Model System for End-to-End Projects Orchestrates 19 AI models in parallel, with Opus matching each task to the best-suited model. Can research, design, code, deploy and manage projects end-to-end. Available to Max subscribers with 10,000 credits/month.
Claude Cowork Update — Google Workspace, DocuSign, and More Anthropic adds connectors for Google Workspace, DocuSign, Apollo, Clay, Outreach, SimilarWeb, MSCI, FactSet, WordPress, and Harvey. Claude can now work across Excel and PowerPoint seamlessly. Plugin marketplace for enterprise admins.

Quick hits — benchmarks, policy, and local model tools.

Update: Hegseth Sets Friday Deadline for Anthropic on AI Safeguards Following yesterday's coverage, Defense Secretary Hegseth has given Anthropic a specific ultimatum: strip Claude's safety guardrails by Friday or face government retaliation.
Update: Qwen 3.5 Tested on 70 Real Repos — Craters on Hard Tasks APEX benchmark tested all Qwen 3.5 models and Codex 5.3 on 70 real codebases. Qwen 3.5 significantly underperforms on hard coding tasks despite strong benchmark numbers elsewhere.
Google's Aletheia Math Agent Solved 6/10 FirstProof Problems Google submitted Aletheia's answers before official release. All prompts and model answers posted on GitHub. A significant step in AI mathematical reasoning.
LM Studio LM Link — Use Your Local Models Remotely Shadow-dropped feature lets you access models on your workstation from your laptop via Tailscale. End-to-end encrypted, integrated in the GUI. Free for up to 2 users, 5 devices each.
From Reddit

Google expands Flow into a creative suite, a solid local image manager ships cross-platform, and the censorship debate heats up between Flux and Qwen Image.

Google Labs Expands Flow Into a Full AI Creative Studio Google Flow gets a major expansion with new creative tools. Building on the Veo video model, Flow now offers a full suite for AI-assisted creative production including editing, effects, and multi-modal generation.
Latent Library v1.0.2 — Local Image Manager for AI Generations Formerly AI Toolbox, now rebranded with cross-platform support. Refactored indexing engine with batch processing, universal metadata parsing for ComfyUI/A1111, visual duplicate detection, and offline AI auto-tagging via WD14 ONNX.
BFL Research Shows Qwen Image Is Much More Uncensored Than Flux 2 Black Forest Labs published research comparing content restrictions between Qwen Image and their own Flux 2 model, showing significantly different censorship levels. Community notes Hunyuan Image 3 remains underexplored.
Update: Seedance 2.0 Now Available in CapCut Desktop + API Following the model release and copyright controversy, Seedance 2.0 is now live in CapCut desktop with API access available. Multimodal video from text, images, audio or video with director-level control.
ComfyUI-SAM3 Update — Click-to-Segment in Canvas Interactive segmentation now works directly in the ComfyUI canvas—click on the image and get the mask instantly. Native model loading with bf16 and flash attention support.
From Reddit

A Launchpad replacement, local AI image editing, a Quicksilver-inspired launcher, and Retool quietly kills self-hosted plans.

LaunchOS — The Launchpad Replacement macOS 26 Deserves Apple removed icon rearranging and folders from Launchpad in macOS 26. LaunchOS spent 5 months building a better one with customizable grid, folder management, and full-screen organizer. Built in public on Reddit.
SmartPic — 100% Local AI Image Editing via Finder Upscale 4x, remove backgrounds, erase objects—all running on-device via Apple Neural Engine. Right-click from Finder, batch process. Zero data leaves your Mac. One-time purchase.
Tuna — A Modern, Modal Launcher for macOS Built on Quicksilver's ideas with Fuzzy Mode, Leader Mode, Text Mode, and Talk Mode. 100% native Swift. Local dictation, currency/time conversions. Free with soft limits, one-time unlock. No subscription.
Trace — Informed Disk Space Management for macOS Swiss-made app categorizes storage into eight categories including System Data. Offers genuinely informed assistance for selective deletion when macOS warnings appear.
MirrorMask — Reshape What the Algorithms Believe About You Native macOS app that scrapes your Google ad profile, then browses as handcrafted personas overnight to reshape what algorithms think about you. 100% local, no telemetry. $39 once.
Retool Disables Self-Hosted Pricing — Enterprise Only Now No public announcement, just a silent change in docs. Self-hosted deployment now requires Enterprise plan. Users discovered the change and community is not happy.
Cronicle's Successor Xyops Just Dropped The popular distributed task scheduler Cronicle just announced its successor, Xyops—a reimagined task scheduler and runner with web UI. Open source.