From Product Hunt
The board was wall-to-wall agent infrastructure today — a place to deploy agents like web apps, a card that lets one spend money, a directory for them to discover services, and a Fair Source desktop workspace to keep a human in the loop.
Tencent EdgeOne Makers — deploy AI agents like web apps, in minutes
Tencent's edge platform bundles the parts you'd normally stitch together yourself: agent runtime, sandboxed tools, memory, observability, a model gateway, serverless functions and storage — driven by familiar CLI/Git/CI-CD workflows. Framework-agnostic (Claude SDK, OpenAI SDK, LangGraph, CrewAI) and polyglot (JS + Python), with no model or framework lock-in. The catch is the usual one: it's a Tencent-cloud play.
Stripe.Directory — a discovery layer for you and your agents to find businesses on Stripe
A single searchable index spanning Stripe Apps, Projects.dev providers, and mpp.dev machine-payments endpoints, so a developer — or an AI agent — can find and integrate a service without manually hunting for it. Driven by
stripe directory search with structured JSON output and built-in agent skills to discover, evaluate, and integrate autonomously. A genuine piece of agentic-commerce plumbing rather than another SaaS.
Mindstone Rebel — a Fair Source desktop AI workspace that asks before it acts
A local-first desktop app for agentic work, built on the Claude Agent SDK with standard MCP connectors and approval checks gating any sensitive action. Keeps per-user memory, lets you pick the model (Opus/Sonnet/Haiku), runs locally, and ships its code as Fair Source (free for individuals/small teams, commercial licence for larger orgs). The local-first + human-approval + inspectable-code angle is exactly the control a power user wants from an agent tool.
Buy by Agentcard — order DoorDash from Claude
Issues a virtual debit card to an AI agent so it can actually buy things: prompt it, and Agentcard creates the account, fills in address and payment, places the order, and emails you tracking. One-click integration with Claude Desktop, ChatGPT and OpenClaw, with per-card spending limits and real-time spend tracking. Equal parts genuinely useful and slightly terrifying — the agent now has a wallet.
FUTO Swipe — genuinely open models for on-device swipe typing
A rare fully-open alternative to Google's and Apple's closed glide-typing engines: a layout-agnostic encoder, layout-specific decoder, and a small context LM totalling ~2.5M params that run on-device in milliseconds. FUTO released the weights and a C++ beam-search library on Hugging Face plus a 1M+ anonymized swipe-gesture dataset, all MIT-licensed. Ships in the offline FUTO Keyboard (Android today), but the models are small and reusable anywhere.
Off Autopilot — curated, human-written articles about agentic coding
A deliberately anti-slop publication: hand-written, curated essays on agentic coding for experienced developers, as a counterpoint to the flood of AI-generated how-tos. It's a newsletter, not a tool — but on-topic reading if you live in Claude Code and Codex all day.
Liner Developer Platform — build search agents with cheap web search
From the AI research-assistant company: six APIs across a Tools tier (Search, Visualization) and an Agent tier (Quick Answer, Search Agent, Deep Research, Visual Answer) that return cited, source-grounded answers. The hook is web search at $1 per 1,000 requests for grounding RAG/search agents — cheap, though it's a crowded field (Exa, Parallel, Brave, Firecrawl) and "10x cheaper" is marketing.
From Reddit
Qwen flips the agent script with a model that predicts environments instead of acting in them, OpenAI tapes out its first custom chip, Claude moves into Slack as a shared teammate, and Europe and Baidu both ship open — plus a memory-cutting optimizer and a court that's warmed to abliteration.
Instead of acting as an agent, AgentWorld is trained to predict what an environment returns after an agent's action — across seven domains: MCP/tool-calling, search, terminal, software engineering, Android, web, and OS GUI. That makes it a substrate for training and offline-evaluating agents and for generating synthetic trajectories without hammering real systems.
The family ships as a 35B-A3B MoE (~3B active) plus a 397B-A17B variant, both open-weight on Hugging Face. On its own AgentWorld benchmark the little 35B-A3B scores 56.39 overall — beating DeepSeek-V4-Pro (52.97), GLM-5.1 (51.31), Kimi K2.6 (53.42), and even Qwen's own 397B (54.74) — with standout SWE (65.63) and OS (65.92) numbers.
OpenAI tapes out "Jalapeño," its first custom inference chip, with Broadcom
A reticle-sized ASIC co-designed with Broadcom specifically for LLM inference (216GB HBM3E, ~7+ TB/s bandwidth), with OpenAI claiming early performance-per-watt "substantially better" than current SOTA. The flex is the schedule: design-to-tape-out in nine months — the fastest high-performance ASIC cycle they're aware of — partly because OpenAI used its own models to accelerate the design. Runs at scale late 2026, part of a 10GW accelerator commitment, and a clear shot at reducing Nvidia dependence.
Anthropic launches Claude Tag — a shared Claude teammate that lives in a Slack channel
Shifts the unit of collaboration from the private chat to the team channel: anyone types @Claude, delegates a task, and it works through it in stages over hours or days, posting threaded updates while the team does other things. Each channel gets one shared Claude with persistent memory, an org identity, and admin-defined tool access. Anthropic says its own product team generates 65% of its code with an internal version — including most of the code that built Claude Tag. Beta for Enterprise/Team; the old Claude-in-Slack integration retires Aug 3.
Baidu Unlimited-OCR — a 3.3B model that parses dozens of pages in one pass
An MIT-licensed multilingual OCR model that does full-document parsing (single images, multi-page docs, PDFs) with 32K output, rather than cropped-region OCR. The trick is Reference Sliding Window Attention (R-SWA): the image tokens stay fully visible to every generated token while generated text only attends to a sliding window — avoiding the KV-cache blowup that makes page 20 far costlier than page 1 in end-to-end OCR. Ships Transformers inference and SGLang serving with OpenAI-compatible streaming (arXiv 2606.23050).
The EU is funding its own open-source 400B+ frontier model
The European Commission named the EUROPA consortium (led by Italy's Domyn) winner of its Frontier AI Grand Challenge, with a plan for an open-source model over 400B parameters covering all 24 official EU languages, trained on European public supercomputers. The prize is compute, not cash — up to 2.5% of total EuroHPC capacity for a year on AI-optimized machines — which directly targets Europe's GPU-access bottleneck. No timeline or training-cost figures yet; Domyn already ships a closed 260B model and an open 10B one.
Qualcomm buys Modular for ~$3.9B — Mojo and MAX go to a chipmaker
An all-stock deal for the startup behind the Mojo language and the MAX platform, whose whole pitch is running AI efficiently across Nvidia, AMD, Intel and other hardware. It's Qualcomm buying its way into data-center AI software and taking a direct swing at Nvidia's CUDA moat. Closes H2 2026 pending approvals; open question is what "open-source Mojo" means under a hardware vendor.
Gefen — a drop-in AdamW replacement claiming ~8x less optimizer memory
A new optimizer pitched as a straight swap for AdamW that cuts optimizer-state memory roughly 8x during training, with code on GitHub (ndvbd/Gefen). With DDR5/HBM prices spiking, anything that shrinks the training memory footprint matters to the independent-training crowd — though the 8x headline deserves benchmark scrutiny before you bet a run on it.
A Swiss high court is evaluating abliteration to counter LLM over-refusal
A paper on over-alignment in multilingual criminal-law settings reports that the Swiss Federal Supreme Court is evaluating the open-source Heretic abliteration project — reaching a favorable conclusion — because off-the-shelf LLMs refuse legitimate criminal-law queries. A notable institutional case of decensoring being treated as a legitimate alignment-mitigation technique rather than a jailbreak.
From Reddit
The community is already sanding down Krea 2's rough edges with a better decoder, a solo dev turned a single image into a playable game on one GPU, and there's a tidy ComfyUI trick for Ideogram 4 brand palettes.
Fixing Krea 2's weak VAE with NVIDIA's PiD pixel-space decoder
Krea 2 ships with the Qwen-Image VAE, which over-smooths output and crushes contrast and fine detail. This workflow swaps the VAE decode step for NVIDIA's PiD (Pixel Diffusion) decoder, decoding 1MP latents directly to 4MP in pixel space for visibly sharper detail and color. Full ComfyUI workflow JSON is posted (Gemma-2-2B text encoder + a 4-step PiD Qwen-Image decoder via Comfy-Org/PixelDiT); caveats are ~15GB VRAM and turning SageAttention off.
A ~0.5B world model that turns any image into a real-time playable game, locally
A solo researcher's from-scratch causal transformer takes a single image and generates a keyboard-controllable, real-time playable "game" on consumer hardware (an RTX 5090) — using LLM-style KV-caching for autoregressive frame generation rather than a giant video model. It's still rough (motion artifacts, context drift over time), but it's a striking local-first counterpoint to the datacenter-scale game-world models, and the architecture writeup is the real draw.
ComfyUI nodes for structured Ideogram 4 color palettes
Custom nodes that let you pick a structured color palette for Ideogram 4 instead of hand-typing hex values into its JSON prompt box, then drive one prompt across four reference modes — logo, monochrome, painting, photograph — from a single palette. Handy for generating on-brand asset variations with consistent color. Low engagement, but the technique is the point.
From Reddit
A strong local-first Mac haul today — a Finder command bar, an rsync GUI, an open Finder replacement, an on-device AI workspace with an evidence ledger, and a native traffic proxy that even speaks MCP — plus a self-hosted job-search cockpit and a 40KB data grid.
Substage 1.0 — a natural-language command bar bolted onto Finder
A command prompt that attaches to Finder windows and runs plain-English actions on the files you've selected. v1.0 adds Rules (customizable system-prompt presets), an Ideas & Actions panel of hundreds of built-in commands, and a redesigned UI. Frequent operations like format conversions run instantly with zero latency, an AI model handles the complex requests, and a safety model pre-parses commands first — CLI power without memorizing flags.
JobOps — a self-hosted cockpit for the job search (that deliberately won't auto-apply)
An open-source job-search command center for searching, tracking, and tailoring applications, pointedly leaving out auto-apply so you stay in control of what goes out. A privacy-respecting alternative to SaaS job trackers; the strong upvote count signals real demand in the self-hosted community.
DeltaSync — a simple, native rsync GUI for the Mac
A clean native front-end over rsync's reliable incremental copy/mirror behavior, giving you predictable local folder backups and mirroring without dropping into a shell or writing a script. A good fit if you trust rsync but don't want to babysit the command line.
Rascal — a free, open-source, customization-first Finder alternative
A file manager aimed squarely at people who find Finder limiting and want to structure things their own way, with deep configurability front and center. Free and open-source (github.com/chang-07/rascal), which makes it an easy thing to just try.
Canto — a fully local AI workspace for Mac with an evidence ledger
Combines chat, a cited Research mode, an AI canvas, code notebooks (Python/JS/TS), and a vault organizer — all on-device. Research plans a run, reads across your vault, attachments and the web, then produces a document with an "Evidence Ledger" listing every claim with Support / Against / Confidence columns and inline citations. A privacy-first alternative to feeding personal notes into cloud chatbots, with a one-time upgrade for unlimited AI.
Ottex — local-first dictation that goes voice-to-finished-result
Beyond voice-to-text: Ottex detects the app or site you're in (Slack, Gmail, Linear, Notion, a CRM) and applies your own custom rules to format and polish the output in one shot, with you approving every result before it's pasted. A Wispr Flow + Granola alternative that lets you bring your own API key or run Whisper fully local for free.
Rockxy — an open-source native macOS HTTP debugging proxy that speaks MCP
A SwiftUI/AppKit (not Electron, not Java) Charles/Proxyman alternative built on SwiftNIO: intercept HTTPS, inspect APIs, set breakpoints, Map Local/Remote, replay requests, and debug WebSocket and GraphQL traffic from Mac apps, CLIs, and the iOS simulator. It ships an MCP server so Claude Desktop can read captured traffic directly. AGPL-3.0 open-core (HTTPS interception, breakpoints, scripting free); Pro from $39.
Snick — mechanical scroll sounds and trackpad haptics for your Mac
A tiny menu-bar app that adds satisfying mechanical scroll sounds plus real Force Touch haptics, synced in lockstep with every scroll notch — seven hand-tuned voices, speed-aware click density, no installer or account. Pure tactile polish; $6.99 one-time (launch discount), macOS 14+.
LyteNyte Grid — a 40KB React data grid that renders millions of rows at 60fps
A zero-dependency, headless-or-styled React grid with 150+ enterprise-grade features — sorting, filtering, grouping, aggregation, range selection — in roughly 40KB. The core edition is free and open-source; PRO adds server-side data, pivoting, and tree data. A genuinely lightweight alternative to AG Grid.