Report by The Next New Thing.
Scroll for the report ↓
Give your AI coding assistant a way to see and press the buttons in apps on your computer, like Finder or Slack, so it can finish tasks there while you keep working.
It pulls each app's accessibility tree as text with a reference for every button, so the agent acts on exact elements instead of guessing pixels.
Supports macOS 13+, Windows 10 1809+ and Ubuntu 24.04+, each through that system's own accessibility interface.
Version 1.2.0 (Nov 5) adds an MCP server (a standard plug-in port for AI tools), so agents can call every command as a tool.
vercel-labs/agent-browserBrowser automation CLI for AI agents
vercel-labs/skillsThe open agent skills tool - npx skills
vercel-labs/agent-skillsVercel's official collection of agent skills
Imagine mouthing a word without making a sound and your Mac typing it. This test app reads the tiny head and jaw movements your AirPods sense and guesses which of your practiced words you said.
View on GitHub →
It matches your movements against word samples you recorded yourself, using a small fixed vocabulary instead of an AI model.
An Evaluate button compares accuracy against shuffled labels 500 times and reports a p-value, so you see whether it beats chance.
Sessions save motion data, mic audio and labels so you can compare normal speech, whispering and silent mouthing.
joeblau/FlappyBlockFlappy Block No Mercy: You lose the game and you have to quit the app.
joeblau/apple-bundle-identifiersApple iOS 17.0 Native App Bundle Identifiers
joeblau/sample-elevator-control-systemJAVA Elevator Control System (3 Hour Challenge)
In Subvocal's Show HN (830 points, 390 comments), one commenter linked Subvox as an open-source version that is 'far from polished.' Another asked about 'the AirPods tweet making the rounds (i.e. the subvocal killer)' and how much custom hardware is really needed. Subvocal's founder says its glasses use RF, which differs from Subvox's motion-sensor approach.
Open the Show HN thread →Think of it as a toy pointer for your Mac screen. A little hand on a stick follows your mouse, taps when you click and draws when you drag, which helps on calls and screen recordings.
A click plays a tap animation and sound, and nothing underneath gets clicked, so you can point safely during a demo.
Hold and drag to sketch over anything; Control+Option+X clears it and Esc turns everything off.
A small toolbar lets you switch pointers or mute the tap sound; the app updates itself.
nilbuild/developer-roadmapInteractive roadmaps, guides and other educational content to help developers grow in their careers.
nilbuild/design-patterns-for-humansAn ultra-simplified explanation to design patterns
nilbuild/driver.jsA lightweight, dependency-free JavaScript library for guiding user focus across the page.
Imagine your coding assistant keeping a notebook. While you work it writes down what it fixed and decided, then hands the useful notes back when you start a new session, and you stop re-explaining your project.
View on GitHub →
It records what your agent does, compresses it into short notes with AI, and feeds the relevant ones into your next session on its own.
The project names Claude Code, Codex, Gemini, OpenCode, Copilot, OpenClaw, T3 Code and, since Nov 2, Devin.
The local engine is free and open source. CMEM Pro adds hosted note-taking and sync across machines: 30-day free trial, then $30 a month.
thedotmack/mcp-client-cliCommand-line interface for any Model Context Protocol (MCP) server.
thedotmack/aims⚡ AI Messenger Service - Watch AI bots communicate in real time
thedotmack/mackeroni-skillsA collection of specialized agent skills for Claude Code and Gemini CLI.
On Oct 29 a user showed that the setup page uploaded "the full text of those prompts" from the last 30 days, with project names and folder paths, to cmem.ai, and that "setup offers no way to stop it." cmem.ai told him the prompts went to a DeepSeek model through OpenRouter without zero-retention settings. v15.8.3 (Nov 2) now asks first, but the issue is still open while cmem.ai checks its OpenRouter settings.
Open issue #5176 on GitHub →Give your coding agent a request like a booking page or a shop. It offers three finished visual styles, builds the page around what visitors came to do, then checks it in a browser for the usual signs of machine-made design.
View on GitHub →
Each theme is a full look (paper, type, colour, texture) with a sample page. It deals three that fit your prompt, or say custom for none.
It renders the page at eight screen widths and flags tells like one coloured word in a headline, then a fresh reviewer gets the last word.
/hallmark redesign rebuilds an existing page and keeps its content and links. /hallmark audit only reports what looks off.
Nutlope/roomGPTUpload a photo of your room to generate your dream room with AI.
Nutlope/aicommitsA CLI that writes your git commit messages for you with AI
Nutlope/logocreatorA free + OSS logo generator powered by Flux on Together AI
Instead of opening a separate app for each coding assistant, you can run Claude Code, Codex, Cursor and others from one desktop, web or phone app, using the subscriptions you already pay for.
View on GitHub →
v0.1.7 (Nov 4) adds Kiro as an early-access agent, tasks scheduled for one date and time, and thread links you can share.
v0.1.6 (Oct 31) lets you ask a side question without breaking the main thread, pin projects, and use 1Password for logins.
MIT licensed. The project says there is no way to pay it; you use the agent subscriptions already on your computer.
pingdotgg/uploadthingFile uploads for modern web devs
pingdotgg/ts-rustAn experimental Rust port of the TypeScript 7 compiler (tsc)
pingdotgg/zactNothing to see here
Theo's PR, merged Nov 3, waits until you are idle for 30 seconds, then shows a "Star on GitHub" button that stars the repo "through the server's GitHub credential" instead of opening a browser tab. It follows the Oct 31 prompt that Theo said "10x'd the number of stars per day," so this week's 14,580-star jump partly measures the prompt.
Open PR #29922 on GitHub →Monid is the OpenRouter for agent tools. Instead of a separate subscription for every API your agent touches, scraping, enrichment, search, social data, it's one key into 1,800+ tools and you pay per call. Same layer OpenRouter built for models, the one Stripe just bought for $7 billion, except for the tools agents actually need.
Monid
Think of it as a design coach living inside your coding agent. You give it named moves like bolder, quieter or polish, and it scans your pages for the telltale habits that make AI-built sites look alike.
View on GitHub →
Commands like bolder, quieter and polish each load a playbook for one job, aimed at one target such as your pricing cards.
Fixed rules flag tells like purple gradients and nested cards, plus cramped padding and skipped headings. No AI model or API key needed.
Live mode overlays your running site on your own computer. Pick an element, try variants in place, and the one you keep is written to your code.
On November 2 a user evaluating Impeccable for a client found that live mode's Apply button can launch claude or codex with approvals and sandboxing turned off, and that this was only written down in an internal document and the Rust source. They asked only for disclosure, not a design change. The project closed it the same day with a new warning in its main docs and the IMPECCABLE_LIVE_COPY_AGENT switch to keep Apply in your chat or turn it off.
Open issue #1274 on GitHub →Give your coding agent a manager's playbook. It asks what you want and why, writes a plan you approve, builds with tests first, and checks its own work before calling anything finished.
View on GitHub →
Jesse Vincent, who made Superpowers, applied the principles he learned from Cialdini’s book Influence to his skills. A Wharton study co-written by Dan Shapiro and Cialdini then tested them on GPT-4o mini: across 28,000 conversations, persuasion more than doubled compliance, from 33% to 72%.
Rules are orders, not tips: “YOU MUST,” “No exceptions.” If a skill applies, “YOU DO NOT HAVE A CHOICE.”
The agent must announce the skill it is using, pick A, B or C, and turn every checklist into todos.
Steps are tied to the moment: “IMMEDIATELY request code review before proceeding.”
Norms stated as facts: “Checklists without todo tracking = steps get skipped. Every time.”
It lists the excuses agents make, like “I know what that means,” and answers each one.
Skills are tested on agents under fake pressure, like production down at $5k a minute, and made stronger after every failure.
Version 7.0.2 opens by asking why you want something, asks only questions that change what gets built, and offers options with a pick.
It can hand every planned task to a new sub-agent with a review after each, or run the plan in one session to spend less.
The project lists setup for Codex, Cursor, Gemini CLI, GitHub Copilot CLI, Grok Build CLI, OpenCode, Pi, Hermes Agent and Muse.
obra/Youtube2WebpageI learn much better from text than from videos
obra/superpowers-marketplaceCurated Claude Code plugin marketplace
obra/superpowers-skillsCommunity-editable skills for Claude Code's superpowers plugin
After 7.0.0 on October 10, long-time users said the rebuilt brainstorming asked pointless questions the code already answered. The team shipped a retune in 7.0.1, then rewrote it again in pull request #2644, merged November 2 and released as 7.0.2 on November 3, thanking everyone on #2530 who sent transcripts. The issue is still open with 30 comments, and the release asks people to keep reporting bad questions there.
Open issue #2530 on GitHub →Think of it as a shelf of ready-made job recipes from the company that makes Claude. Each folder teaches Claude one task, like filling in a PDF, building a Slack GIF, or writing a new skill of your own.
View on GitHub →
The Word, PDF, PowerPoint and Excel skills that power Claude's document features are here to read, under a restrictive license.
skill-creator helps you build your own skill, and a plain template folder shows the two fields every skill needs.
Add the repo as a plugin marketplace, then install document-skills or example-skills. Paid Claude.ai plans already include these.
anthropics/claude-codeClaude Code is an agentic coding tool that lives in your terminal, understands your codebase, and helps you code faster by executing routine tasks, explaining complex code, and handling git workflows - all through natural language commands.
anthropics/claude-cookbooksA collection of notebooks/recipes showcasing some fun and effective ways of using Claude.
anthropics/claude-plugins-officialOfficial, Anthropic-managed directory of high quality Claude Code Plugins.
Describe what you want in Claude or ChatGPT, and your AI builds the Zap and turns it on.
The predictable steps run as plain code, and AI only handles the parts that need judgment.
Zapier’s example: a brief before a big call dropped from $1.12 a run to 12 cents.
When a Zap breaks, it can figure out how to recover on its own.
It’s in beta now, with a waitlist open.
zapier.com/s/ngz-join-wait-listWhen you run several AI coding agents on one project, it gives each one a separate copy of the code to work in, then helps you merge the good work back and clean up the rest.
View on GitHub →
wt switch, wt list and wt remove create, check and delete a separate work folder by branch name, in place of longer plain git commands.
wt switch -c -x claude feat makes a new branch and folder and launches Claude Code inside it.
wt merge squashes, rebases and merges the branch into main, can draft the commit message with an AI model, then removes the folder.
Give your coding agent these ready-made design checklists and it reviews a screen for color, spacing, wording and accessibility, then fixes what it finds. One of them even draws a labeled diagram of any object.
View on GitHub →
better-interface checks a screen, flow or whole interface across six areas, including accessibility and writing, and returns one prioritized list.
interface-review checks a branch, pull request or uncommitted change for interface problems it introduced or made worse.
teardown makes a technical illustration of an object that you can label, color and download as SVG, PNG or PDF.
jakubkrehel/make-interfaces-feel-betterAn agent skill that helps make your interface feel better.
jakubkrehel/loadingA lightweight library full of beautiful loading indicators for React.
jakubkrehel/oklch-skillAn agent skill that helps you work with OKLCH colors.
Think of it as a workbench for AI coding. You open a project, give each task its own separate copy of the code, chat with the agent, check its changes and send them to GitHub, all in one window.
View on GitHub →
Each task gets its own git worktree, a separate branch and folder, so several agents can work at once without touching your main copy.
The agent keeps a linked set of SPEC.md files beside the code, so what the project is meant to do carries over between sessions.
Sign in with JetBrains AI, connect a provider through pi, or point it at a model on your own machine such as Ollama or LM Studio.
JetBrains/kotlinThe Kotlin Programming Language.
JetBrains/intellij-communityIntelliJ IDEA & IntelliJ Platform
JetBrains/compose-multiplatformCompose Multiplatform, a modern UI framework for Kotlin that makes building performant and beautiful user interfaces easy and enjoyable.
Instead of the plain terminal that came with your computer (the window where you type commands), you get a quicker app that feels native on Mac and Linux, with tabs, split panes and pictures right in the window.
View on GitHub →
1.4 quietly compresses scrollback you are not looking at, which cuts memory use, according to its release notes.
It now covers 100% of the Kitty image standard, animation included, which the project says no terminal outside Kitty had done.
The core is a C and Zig library anyone can use to build a terminal or embed one in their own app.
ghostty-org/ghostlingA minimum viable terminal emulator built on top of the libghostty C API. Ex minimo, infinita nascuntur. 👻🐣
ghostty-org/discord-botGhostty Discord bot
ghostty-org/zig-gobjectGObject bindings for Zig using GObject introspection, including specifically the libraries Ghostty requires.
On Apr 28, 2026 Mitchell Hashimoto wrote that GitHub outages were blocking his work almost every day and that Ghostty would move elsewhere, keeping a read-only mirror. Six months later the project still runs on GitHub, and an HN commenter on the 1.4 thread asked what had changed.
Read Mitchell's post →A local plugin for Claude Code and Codex CLI that folds searches and edits that normally take several tool calls into one MCP call (a plug-in tool connection) with compact results. It also searches your past sessions, recovers decisions after the context gets compacted, and runs a code review from several angles. No account, telemetry or subscription.
View on GitHub →Alessio is a developer and a viewer of the show. He built TokenSaver to cut the tokens Claude Code and Codex burn on back-and-forth tool calls, and kept it fully local: no account, no telemetry, no subscription.
Building something? Send it to hi@TheNextNewThing.ai
Two skills for Claude Code and Codex. /skeleton lists and writes the basics an app needs before its first feature, such as tests on every push, a .env.example, a /health route, backups with a rehearsed restore, a rollback plan and security headers. /skeptic-review audits an app that is already running and points to the file and line for each answer.
View on GitHub →Kristjan, from Estonia, audited his own app after three weeks of daily use and found 22 issues, 15 of them preventable on day zero. So he wrote down the 13 things he wishes he had set up first.
Building something? Send it to hi@TheNextNewThing.ai
