From Product Hunt

A very agent-shaped launch day: MCP bridges, local memory for coding agents, and a GLM-5.2 harness — plus a couple of native Mac apps worth a look.

Macuse — Give your AI superpowers on macOS Native Mac app that connects Claude, Codex, Cursor, Raycast and any MCP client to your Mac apps — Calendar, Mail, Notes, Reminders, Messages, Contacts — plus Computer Use to drive apps without APIs. Local-first, revocable per-app permissions, free.
▲ 111 · macuse.app
Flowly — Open-source personal AI agent Now fully open source (Apache-2.0). Runs on your own hardware (macOS/Win/Linux/iPhone/Android) with your own API keys, summoned from the menubar, notch, or a global hotkey. Keeps a private, self-correcting memory of your world, plus browser automation and cross-device sync.
▲ 104 · useflowlyapp.com
scritty — Shared, searchable memory for every AI coding agent A terminal that passively captures every CLI agent's conversation (Claude, Codex, Copilot, Antigravity, Ollama), indexes it into one local corpus (hybrid keyword + vector, fully offline), and serves it back to agents over MCP and to you over the CLI. Session sync across desktop, browser and mobile. $19.99/mo, local-first.
▲ 115 · scritty.dev
Context.dev — One API to scrape, enrich and extract the internet A single web-context API for agents: scrape any URL to clean markdown, crawl sitemaps, extract structured data into your own schema, capture screenshots, pull brand data and company enrichment. TS/Python/Ruby SDKs and agent-native (paste one line and your coding agent wires it in). YC-backed, free tier, #1 on PH today.
▲ 587 · context.dev
ZCode — The official harness for GLM-5.2 Z.ai's own agentic dev environment for GLM-5.2, built for long-running coding: stable context across file edits, terminal and browser state, and Git review, plus a mobile Remote, bot control and BYOK. macOS, Windows and Linux apps.
▲ 51 · zcode.z.ai
PieterPost MCP — Connect your AI agent to postal mail An MCP server that lets ChatGPT, Claude or any MCP client prepare real letters and postcards, use Mailbook contacts, attach files or postcard images, create checkout links and track orders. The MCP is free; you only pay per item actually printed and mailed.
▲ 127 · pieterpost.com
Retrace — Debug AI agents by replaying and forking runs Record, replay, fork and share agent executions. See every LLM call, tool invocation and error, then fork a run at any point to test a fix without re-executing side effects like payments, emails or API calls. SDK on GitHub, free tier of 1,000 traces/mo.
▲ 92 · retraceai.tech
Quick Sub 2 — Creative video subtitling for Mac Native SwiftUI subtitling with direct canvas control: drag subtitle objects on the video, independent text/container/rotation styling, batch styles, a timeline that scales 0.1x–10x, native undo/redo and .qsub2 project files. Free.
▲ 87 · apps.apple.com
From Reddit

A loaded day: Anthropic walks into pharma, Fable 5 comes back with a controversial safety leash, OpenAI floats giving Washington a stake, and the local scene keeps compressing frontier models onto single GPUs.

Fable 5 is back — but a new safety classifier is quietly downgrading it After the June 12 export-control takedown, Fable 5 and Mythos 5 returned July 1 with a new classifier that catches the reported jailbreak 99%+ of the time and silently reroutes flagged requests to Opus 4.8. BridgeMind's rerun of BridgeBench shows big drops vs the June build (debugging 86.2 → 25.9, refactoring 73.6 → 38.4); the claim is over-triggering on normal coding, not changed weights. Pre-nerf, Fable 5 had scored 16.10% on the Remote Labor Automation index, double the next-best model.
OpenAI reportedly floats a 5% stake for the US government Per FT/CNBC, OpenAI proposed giving the Trump administration roughly a 5% stake to ease Washington pressure — the same pressure that staggered GPT-5.6's launch to a handful of government-approved orgs. It lands ahead of OpenAI's confidential S-1 and amid the "why can Anthropic ship Fable but OpenAI can't ship GPT-5.6" grievance.
DeepSeek V4 Flash running at full 1M context on a single RTX 5090 The model demanded ~256GB VRAM at 1M context because llama.cpp lacked support for the DSA lightning indexer. A community dev wired in the upstream PR and patched a CUDA kernel: at 256K context the compute buffer dropped from ~67 GiB (OOM) to 3.2 GiB, prefill jumped 56 → 263 t/s, and 1M context went from impossible to working (3.75 GiB at ubatch 768) on one 5090. A landmark for running a frontier MoE locally.

Quick hits — open models, tooling and research.

poolside Laguna-XS-2.1 — a competitive US open model Small open-weight model landing very close to Qwen 3.6 35B-A3B; llama.cpp support merged same day.
Kimi K2.7 Code goes GA in GitHub Copilot The first open-weight model offered as a selectable option in Copilot's model picker.
"Is One Layer Enough?" New paper claims training a single transformer layer can match full-parameter RL training on some tasks.
From Reddit

Practical control was the theme: a LoRA that finally aims the sun, an open infographic model that rivals Ideogram, plus the Krea 2 workflow ecosystem maturing.

Flux 2 Klein LoRA for precise sun direction A LoRA that precisely controls the sun's direction and elevation on any exterior image — instead of begging the prompt with "sunlight from the left" forty times. Weights on Hugging Face; a v2 adding hardness, color and intensity controls is in progress. Genuinely useful for product, architecture and website graphics where consistent lighting matters.
SenseNova-U1-8B-MoT-Infographic — open SOTA for infographics + editing A Mixture-of-Transformers image model tuned for dense infographic design and image editing — the only thing besides Ideogram 4 that gets close on infographics, and it's Apache-2.0. Handles 100+ styles/layouts with strong small-text rendering (arXiv-style dense pages). Not GGUF yet, but you can wrap it in a FastAPI OpenAI-compatible image endpoint.
A 60-hour-tested Krea 2 realism workflow + settings guide A deep ComfyUI workflow and settings write-up for high-quality, facially-expressive realistic images out of Krea 2 (the open-weights model shadow-dropped last month), including an optional full-turbo variant and sampler/LoRA picks. The kind of "how", not just results, that's worth saving.
Krea2 LoRA Stripper — shrink LoRAs ~90% Open-source script that strips the DIT/UNET blocks and keeps only text-conditioning layers, since the base model already knows the visuals.
VR-Outpaint 1.0 — turn flat video into 360° An LTX 2.3 IC-LoRA that outpaints a flat clip into a full equirectangular sphere via an inverse gnomonic projection trick; ComfyUI workflow included.
From Reddit

A homelab timeline, a Finder that finally gets a "New File", a self-hosted AI gateway, and — at last — an official Claude Desktop for Linux.

Tempo — one Mac timeline for your whole homelab A ~13MB native app that pulls every signal from your setup (Home Assistant, Uptime Kuma, Kopia backups, UniFi, CI/CD, custom scripts) onto one live timeline — a host dropped and came back, a backup finished, a script's output — newest first. Every event carries buttons you define (SSH the host, run a script, restart a container, copy a payload value). Local-first, no cloud, LAN ingestion server.
PowerClick — make Finder's right-click actually useful Finally adds "New File" to Finder's right-click (text/code/web/data + custom templates), plus open-in-VS-Code, batch rename, move, copy path/info and clipboard handling — without leaving Finder. Works inside iCloud/Dropbox/Drive/OneDrive. $5.99 one-time, universal, macOS 13+.
Facet — local, open-source photo culling engine MIT-licensed photo scoring and culling with a self-hosted web gallery, no cloud. Scores and helps you cull large photo sets, then browse the keepers locally — a good fit for a privacy-first, self-hosted setup.
OmniRoute — self-hosted AI gateway, 237 providers One OpenAI-compatible endpoint across 237 providers (90+ free), with automatic fallback and a 10-engine token-compression pipeline. MIT, ~9.8K stars, BYOK and self-hosted — squarely in "own your routing across every model" territory.

Quick hits — more tools and one official release.

Claude Desktop lands on Linux (official beta) Chat, Cowork and Code for Ubuntu 22.04+/Debian 12+ via Anthropic's signed apt repo. Computer Use and voice are disabled for now.
TaskView 1.48.7 — self-hosted project management Big release: new dashboard, recurring tasks, time tracking, sprint planning and heavy field/analytics customization. Own your data.
Speech Studio — local voice cloning for macOS Open-source (Apache-2.0), built on the soniqo speech-swift toolkit (MLX + CoreML). Clones a voice fully on-device on Apple Silicon.
Trovelo — private, offline trip planner for Mac iCloud-only itineraries, maps, multi-currency budgets and packing lists — no account, no servers. Free tier; Pro is a $6.99 one-time.
justsaydone — trim Claude Code's verbose output Open-source tool that claims up to 95% shorter replies to cut output-token spend on long agentic sessions.
From Twitter

On Twitter today, Fable 5's return turned into an orchestration playbook, open weights collected receipts, agents moved into the browser, and Etched walked out of stealth.