The model stopped being the product today.

Moonshot's desktop agent for Windows and macOS (Apple silicon), pitched as "the AI Desktop for Knowledge Work" and built like one: it reads and edits files in your local workspace behind an explicit authorization prompt before anything is modified, overwritten, or executed; a built-in cron engine schedules agent calls and Python or shell jobs daily, hourly, or conditionally; WebBridge drives a real browser — clicks, scrolls, extraction — and an Agent Swarm mode coordinates specialized sub-agents on long tasks. The lineage is the point: Kimi K3 crossed our pool Friday as a model launch — 2.8 trillion parameters, "the world's first open 3T-class model," full weights promised by July 27 — and today the same vendor shipped the thing it actually charges for. The pre-integrated market data (A-shares, Hong Kong, US equities) says who they think signs up first.

Reach for it when you want Friday's LM Studio Bionic pattern — the vendor's own all-in-one agent desktop — from the open-weights side of the argument. Delete the pile of half-configured schedulers and browser-automation scripts pretending to be one assistant. Tradeoff: the app is closed and membership-priced, and it is a Beijing-based lab's agent with a browser, a cron daemon, and authorization-gated access to your local files — the gate is stated plainly, but the data-residency call is yours to make before it touches client work.

A native SwiftUI macOS app that bundles an mlx-vlm server and wraps the whole local-AI loop around it: chat with vision and reasoning output, a model library that discovers what's already in your Hugging Face cache, a performance dashboard (time to first token, decode speed, token volume), and localhost endpoints speaking both the OpenAI and Anthropic dialects. The detail that matters here: an integrations pane configures and launches Codex, Claude Code, Pi, Hermes, and OpenCode against models Nativ serves — pointing a coding agent at a local open model stops being a proxy-chain weekend project. The deep knobs (KV-cache quantization, prefix caching, speculative decoding) live in the UI, not a config file. MIT, 426 stars, created yesterday — by Prince Canuma, author of mlx-vlm itself, the library much of Mac-local vision inference already runs on, which is why a day-old repo gets a pick. Friday's Bionic was the local-model vendor's closed, cloud-metered version of this instinct; four days later, here is the open one.

Reach for it when the open-weights argument gets concrete — running the models the front page was arguing about on hardware you own, with no account anywhere. Delete the tmux pane running a Python inference server and the second one pretending to be its dashboard. Tradeoff: a day old, macOS 26+ and Apple-silicon-only, and the app can't fix the physics — quality and speed ceilings belong to the model and your unified memory, not the wrapper.

Point it at a URL and get the site back as componentized Next.js or Vite code — real components, design tokens, and an editable content model, with hover and focus states, dropdowns, accordions, motion, web fonts, SEO metadata, and responsive layout preserved. The design bet is deterministic extraction instead of AI guessing, and the receipt is byte-stable output: run it twice, get identical code — which the screenshot-to-code cloners structurally cannot promise. MIT and self-hostable, with a free hosted REST API and an MCP server so agents can call it too. From the ion.design team; 588 stars, created three weeks ago; #5 on Product Hunt today at 126 votes.

Reach for it when a rebuild starts from an existing site — yours after a platform migration, or a design you're matching — and the alternative was a model burning tokens to approximate what a parser can just read. Delete the "clone this landing page" prompts and their subtly different output every run. Tradeoff: pointing it at sites you don't own has legal edges the tool won't police for you; "clean" is the maker's adjective, so let the demo judge; and the hosted API is free today with no published price for tomorrow.

One API call returns a structured JSON action manifest for any webpage: resolved CSS and role locators for everything clickable, fillable, and submittable, plus a requires field encoding the dependencies between actions — select a plan before the submit button becomes usable. Under the hood it is Playwright and DOM analysis for structure, with Claude Sonnet inferring the dependency graph; it ships a Python SDK, a LangChain integration, and an MCP server, and the demo needs no signup. This is the 07-10 watch arriving: we said the second-interface pattern needed its schema-from-source layer, the way OpenAPI standardized REST — Manifest is that layer, built from the outside in: the manifest is derived by a third party, not published by the site's owner. It is also the day's cleanest instance of the theme — spend the model once per page, hand every subsequent agent run a cached artifact instead of a re-derivation.

Reach for it when your browser agent keeps burning turns rediscovering the same form's structure, or your selector files break weekly. Delete the hand-maintained locators and the retry loops that paper over them. Tradeoff: requires comes from static DOM analysis, not live probing — the author states plainly that async server-side validation is invisible to it — single-page apps need a fresh call after state changes, and the whole thing is hosted, closed, and pre-revenue from a solo founder: treat the manifest as cache, not gospel.

An open-source, local-first video editor where AI agents work on a real multitrack timeline — clips, transitions, captions, effects — instead of emitting a finished mp4. A built-in agent handles chat-driven edits, external agents connect over MCP (Claude Code and Codex are the named integrations), and every agent-proposed edit queues for your review before it applies. AGPL-3.0, free with editable exports; 179 stars, created six days ago; the maker is explicit that it is an independent alternative to the commercial ChatCut, with no affiliation. On our video thread this is the missing piece: aftr drove After Effects over MCP (07-13), video-use ships you a final.mp4 (07-20 footer) — OpenChatCut's position is that the artifact should stay editable and the model should stay swappable.

Reach for it when "AI video editing" has meant regenerating the whole output to change one cut — a timeline is exactly the artifact that makes iteration cheap. Delete the black-box generators for any video someone will ask you to revise. Tradeoff: six days old in a deep-end category (multitrack editing is where side projects go to be humbled), the AGPL matters if you plan to embed it, and review-before-apply — the 07-02 last-click gate, now in video — is only as good as your patience with the queue.

Off the thread but worth knowing: movie-gen (MIT, created Friday, 23 stars) is the theme in film form — a Claude Code-directed pipeline over Seedance 2.0, Nano Banana Pro, Sonilo, and ElevenLabs via Higgsfield's CLI, with a playbook distilled from about ten films and an 11-minute, ~80-shot worked example; its core cost move is iterating on the animatic — the whole film as stills plus TTS, nearly free — before any Seedance money is spent, and the author posts the bill: roughly $200 and 2.5 hours per movie. tterm (Product Hunt, 77 votes) puts a terminal, an embedded Chromium, and Claude Code in one closed macOS app with no code editor on purpose — you review the agent's work hunk by hunk; free for hobbyists. Diffsmith (73 votes) is a one-time-purchase, fully offline Mac App Store app for line-anchored comments on your agent's uncommitted diff, with the agent's replies arriving over MCP at the exact line — 07-14's human-legible-interface thread, now sold as shrinkwrap. And the day's only GitHub trending row is ai-agent-book (Apache-2.0, 13,274 stars, 4,434 of them today) — Li Bojie's Chinese-language open-source textbook on agent design with per-chapter code, first published September 2025 and re-trending now: the open-weights zeitgeist arriving as pedagogy, in the language of the labs the front page was arguing about.