From Product Hunt

A heavy day for agent infrastructure: tools that wire your own context into Claude Code, give agents a reliable execution layer, and keep work moving after the code is written — plus a few curveballs in hardware and computational tooling.

Framer 3.0 — Agents, Branching, Community, and an all-new design AI Agents now live on the canvas to help you design, write, analyze, and organize sites — and you can wire your own Claude Code or Codex into Framer. Branching lets teams explore ideas safely before publishing, and a new Community lets creators share work and earn.
▲ 418 · framer.com
Swytchcode CLI — Reliable agent access to 2,000+ APIs An execution layer that sits between your AI agent and every API it calls: schema validation, OAuth/keys handled, idempotency so duplicate runs don't double-fire, and policy enforcement. 2,000+ APIs (Stripe, GitHub, Slack, Notion…) out of the box. Run npx swytchcode; works with Claude, Cursor, Codex, and Hermes.
▲ 349 · swytchcode.com
Daemons by Charlie Labs — Always-on AI agents that keep work moving Persistent, role-scoped daemons that pick up the operational debt coding agents create: monitoring PRs, issues, CI, docs, and Sentry errors over time. You define roles in Markdown; they leave reviewable updates in GitHub, Linear, Slack, and Sentry. "Agents create work, daemons do the rest."
▲ 217 · charlielabs.ai
Quartz — AI email client that runs locally on your Mac Turns Gmail into a focused inbox: sorts mail by importance, learns what matters to you, and drafts replies in your own voice. The AI runs entirely on-device (Gemma) so mail stays end-to-end encrypted and never touches an AI provider. Free public beta.
▲ 205 · quartzmail.ai
Deep Work Plan — Turn any repo into an agent harness Not a chat window it forgets — a spec written into the repo: atomic tasks, acceptance criteria, validation gates, resumable state. Long runs survive context resets and any coding agent picks up where the last left off, coding to a plan it can't drift from. Open source, MIT.
▲ 108 · deepworkplan.com
memi — A macOS agent workbench for product design teams A native Mac workbench where Claude, Codex, and Hermes run on your specs, research, and Figma files — an agent harness aimed at design and product work rather than just code.
▲ 100 · memoire.cv
Wolfram Language 15 — Computational language for humans and AI agents A big release across LLM/AI tooling, symbolic and numeric computation, visualization, geometry and graphs, the knowledgebase, PDEs and system modeling, plus core-language and compiler updates — increasingly positioned as a reliable tool-use backend for agents.
▲ 93 · wolfram.com
Snapchat SPECS — A computer in lightweight see-through glasses Snap's AR spectacles pack a real computer into lightweight see-through glasses, putting apps into the world around you so you can learn, work, and play while staying present. Snap's biggest bet yet on hands-free, glanceable computing.
▲ 85 · specs.com
From Reddit

The strangest launch of the day came from Midjourney, not a frontier lab. Elsewhere: a closer GPT-5.6, more local-model momentum, and the uncomfortable arrival of AI in the literal kill chain.

OpenAI's chief scientist calls GPT-5.6 a "meaningful improvement" — could land this month Per The Information, Jakub Pachocki told staff GPT-5.6 is a meaningful step over GPT-5.5 and could ship as early as June. Benchmarks, pricing, and an exact date aren't confirmed yet — but it lines up with the rumored June timeline and the long-running "iris-alpha" codename.
Gemma 4 runs in-browser at 255 tok/s on hand-optimized WebGPU kernels A demo plus open kernels run Gemma 4 E2B fully in the browser via WebGPU, hitting ~255 tok/s on an M4 Max — no server, fully private. A nice footnote: the kernels were optimized with help from Claude Fable 5 before the model was pulled offline.
Update: GLM-5.2 is already landing on local rigs — GGUF quants incoming Following its launch, GLM-5.2 (753B MoE, ~40B active, MIT) is being called "a win for local AI." Unsloth is uploading GGUF quants, and the community is charting setups from 8×H200 down to a 1-bit dynamic quant on a 192GB Mac Studio or a 24GB GPU + 192GB RAM. The prize many want most: distilling its reasoning into 8–70B daily drivers.
AI in the kill chain: Grok reportedly used to strike Iran as Anthropic pulled out The Pentagon revealed it used xAI's Grok for targeting in strikes on Iran even after Anthropic declined to allow Claude for the operation. Anthropic's CEO says the company "doesn't know exactly how" Claude was used in an Iran school strike, and 200+ organizations are now calling for a ban on AI in military kill chains — putting automated targeting squarely in public view.
Update: The Anthropic–Trump standoff over Fable 5, explained After the export-control order that took Fable 5 and Mythos 5 offline, the administration is reportedly pushing Anthropic for "unbreakable" guardrails — which researchers note is effectively impossible, since any classifier can be jailbroken. Anthropic argues the flagged vulnerabilities are simple and surface in other public models too. Talks continue with no resolution yet.
Inflect-Nano — a usable TTS model in just 4.63M parameters An experiment in how small neural TTS can go: 4.63M params, 24 kHz English, runs locally on almost anything — ~17× smaller than Kokoro and nearly 1000× smaller than Fish Audio S2 Pro. Robotic at the edges, but a striking size-to-function baseline for embedded and offline voice.
From Reddit

Less spectacle, more plumbing this week — a unified video-training config, a new open base model, and LoRAs that make Ideogram 4 cheap to run at home.

Boogu Image — a new Apache-licensed text-to-image model (with an edit variant) A fresh open T2I model released under Apache 2.0, already shipping ComfyUI nodes and a separate image-editing variant. Early community tests pit Boogu's edit model against Flux Klein and Qwen-Image-Edit on identical inputs and seeds — worth a look if you want a permissively-licensed base to build on.
Ideogram 4 gets a Turbo LoRA (2–8 steps) and a half-VRAM Differential LoRA Two Ostris releases make your favorite text-image model far cheaper to run locally. The Turbo LoRA produces usable results at 2–8 steps; the separate Differential LoRA replaces the full 9B unconditional model during the unconditional pass to roughly halve VRAM. Trade-off: Turbo doesn't always play nicely with custom character LoRAs.
SenseNova-U1's 8-step Infographic LoRA — fast, text-accurate layouts An 8B model whose distilled LoRA drops infographic generation from 50 steps to 8 while keeping crisp non-blurry text, structured grids/columns, and accurate chart labels (English + Chinese). Runs on consumer GPUs — genuinely practical for business graphics. ComfyUI nodes available.
From Reddit

A strong week for native, open-source Mac utilities — disk analysis, screenshots, drive safety, and the menu bar — plus a self-hosted SSH manager.

Radix — a fast, free, open-source macOS disk analyzer A native SwiftUI disk-space analyzer that scans any folder or volume and renders it as an interactive sunburst chart. Quick Look on any file, scoped or full-tree search, metadata inspection, and a scanning engine the dev claims beats the competition — a free, OSS DaisyDisk alternative that sips RAM.
macshot — an open-source CleanShot X alternative One reviewer's pick after a frustrating hunt for a decent OSS screenshot tool: macshot has screenshot history, lets you reopen and re-edit previous captures, and gets close enough to a CleanShot X workflow to be worth following. Still rough in spots, but the most promising open option around.
Noject — stop accidental ejects of always-plugged drives A tiny native menu-bar app for setups (like a Mac mini with always-on external NVMe) where macOS insists on treating fixed drives as removable. Noject guards those drives from accidental ejects so your active work, backups, and project data stay mounted.
DynamicLake — a Dynamic Island for the Mac, redesigned for Liquid Glass Brings an iOS-style Dynamic Island to macOS, notch or not: Apple Music queue with live FFT waveforms, interactive notifications from iMessage/WhatsApp/Telegram/Slack, a DynaDrop file shelf with offline conversion, plus calendar and timers — now with a Liquid Glass refresh.
Skiff — an open-source, self-hostable Termius alternative A self-hosted SSH connection manager aiming to replace Termius — open-source and early, but already picking up stars.
Pasly — make Cmd+V smarter A Mac utility that upgrades plain paste with smarter, context-aware paste behavior — handy if you live in the clipboard but want more than a basic history.