Hacker News

Article URL: https://www.statnews.com/2026/02/20/nicotine-wellness-startups-productivity-boost-legal-gray-area/

Comments URL: https://news.ycombinator.com/item?id=47111506

Points: 1

# Comments: 0

Article URL: https://s2transfer.xyz

Comments URL: https://news.ycombinator.com/item?id=47111220

Points: 1

# Comments: 1

Article URL: https://nesbitt.io/2026/02/05/git-magic-files.html

Comments URL: https://news.ycombinator.com/item?id=47111218

Points: 1

# Comments: 0

Article URL: https://georggrab.net/content/opus46retrieval.html

Comments URL: https://news.ycombinator.com/item?id=47111213

Points: 1

# Comments: 0

I completely vibe coded this digital Zen garden to have something to do for the 2 minute breaks that happen when you wait for your AI agent. 10k+ lines JS, 5k+ lines CSS and 0 idea how it really works besides main account login logic and stripe integration. Switched from Claude Code to Codex to Gemini and back to Codex which I feel is the most capable cli coding agent right now. Can do 5+ mins of work without going off rails consistently. This project was made just over 2 weeks and I certainly am feeling the AGI when working with SOTA coding tools

Comments URL: https://news.ycombinator.com/item?id=47111194

Points: 1

# Comments: 0

I have a bit of phone anxiety, and have a ton of dread around making phone calls to restaurants, banks, doctors, and so on and on.

I thought: AI could do this with a web form turned into a prompt.

Stack started out simple -> using 11labs for voice + claude + twillio, but it actually got rather complex (even though I tried vibe coding most).

First off, finding the phone numbers quickly is hard. This is done by scraping the web with some basic duckduckgo search and structure with openai calls.

Second, collecting the right information. I’m still struggling a bit with this but the architecture is that: A) user puts in call objective and business name B) if keywords are detected spin up one of the default form categories C) if not, get structured json from gpt-4o-mini and turn into react form

The cost of making a single call spun out of control, but luckily sonnet can handle a lot of the calls and I’m ok paying for twillio.

Ended up taking months to build my week-long project because of course.

It’s still WIP so feel free to email me: galcohavy@ucla.edu with any ideas or issues u ran into. \

Comments URL: https://news.ycombinator.com/item?id=47111179

Points: 1

# Comments: 1

Claude Code asks for permission before running tools (Bash, Write, Edit, etc.). If you're not at your terminal, it just waits. This tool hooks into Claude Code's PermissionRequest hook and sends each prompt as a push notification to your phone via ntfy.sh. Tap Approve or Deny, and Claude continues.

Setup:

npm install -g claude-remote-approver claude-remote-approver setup Then scan the QR code with the ntfy app on your phone and start a new Claude Code session.

How it works: The hook POSTs the permission request to an ntfy topic, then subscribes to a response topic via SSE. When you tap a button on your phone, ntfy delivers the response back. The hook writes {"behavior":"allow"} or {"behavior":"deny"} to stdout and exits.

The topic name is generated with crypto.randomBytes(16) (128 bits), config file is 0600, and unanswered requests auto-deny after 120 seconds.

If you don't want requests going through the public ntfy.sh server, you can self-host ntfy and point the config at your own instance.

Github: https://github.com/yuuichieguchi/claude-remote-approver npm: https://www.npmjs.com/package/claude-remote-approver

Comments URL: https://news.ycombinator.com/item?id=47111171

Points: 1

# Comments: 0

Article URL: https://ghostty-style.vercel.app/

Comments URL: https://news.ycombinator.com/item?id=47111166

Points: 1

# Comments: 0

Article URL: https://arxiv.org/abs/2602.06811

Comments URL: https://news.ycombinator.com/item?id=47111164

Points: 1

# Comments: 0

Finnish is morphologically complex — vowel harmony, six grammatical cases, free word order. AI models get it wrong in the same predictable ways every time.

I built this after noticing that AI-generated Finnish triggers immediate pattern recognition in native speakers: overly formal register, SVO word order (Finnish allows much more variation), missing discourse particles (-han/-hän, -pa/-pä), and excessive nominalization. Before: "Tämä on erittäin merkittävä kehitysaskel, joka tulee vaikuttamaan laajasti alan tulevaisuuteen. On syytä huomata, että kyseinen innovaatio tarjoaa lukuisia mahdollisuuksia eri sidosryhmille." After: "Iso juttu alalle. En ole varma mihin tämä lopulta johtaa, mutta hyötyjiä on – varsinkin ne jotka ovat odottaneet tällaista jo vuosia." Finnish Humanizer is a pattern library — 26 identified patterns — packaged as a Claude Code skill and distributed for 15 platforms (Cursor, Copilot, Windsurf, ChatGPT custom instructions, etc.). It's not an additional model call. It's a checklist of linguistic tells you can apply mechanically. Patterns are grounded in Finnish linguistics research (Kotus — the Research Institute for the Languages of Finland). Would especially value feedback from Finnish speakers on pattern coverage and any false positives.

Comments URL: https://news.ycombinator.com/item?id=47111131

Points: 1

# Comments: 0

Article URL: https://world.hey.com/dhh/wonderful-vi-a1d034d3

Comments URL: https://news.ycombinator.com/item?id=47111112

Points: 1

# Comments: 0

Article URL: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.chatterjeexi.html

Comments URL: https://news.ycombinator.com/item?id=47111092

Points: 1

# Comments: 0

Article URL: https://github.blog/engineering/from-pixels-to-characters-the-engineering-behind-github-copilot-clis-animated-ascii-banner/

Comments URL: https://news.ycombinator.com/item?id=47111082

Points: 1

# Comments: 0

Article URL: https://www.bbc.com/news/articles/c5yj2kzkrj0o

Comments URL: https://news.ycombinator.com/item?id=47111067

Points: 4

# Comments: 0

I built an open-source bot that turns Trello cards into working code using Claude Code.

Drop a task card in a list → Sergio picks it up, explores your codebase, and posts an implementation plan as a comment. Add feedback, move the card back, and iterate. When you're happy, move it to the Development list → Sergio creates a worktree, writes the code, runs tests, and opens a draft PR on GitHub.

It's a tool that can be used by teams of devs and product managers to cover the knowledge gaps between non technical and technical planning.

All triggered by dragging cards. It's basically Claude Code running as an autonomous teammate on a $5/month VM, orchestrated through Trello. The two-user sandbox architecture keeps the AI isolated from secrets and credentials (similar to OpenClaw's approach to secure agentic coding).

The roadmap includes pluggable engine support (OpenCode, Codex) and MCP servers for reading Google Docs, Figma, and Notion directly from cards.

Comments URL: https://news.ycombinator.com/item?id=47111063

Points: 1

# Comments: 0

I built Paragent because I kept context-switching between features.

The idea: describe what you want in plain English, and an agent branches off, writes the code, and opens a PR. You can run 10 at once — each on its own branch.

How it works: - Connect your repo via GitHub App (minimal permissions: contents + PRs) - Describe a feature ("Add Stripe checkout to the pricing page") - Agent plans, writes, runs your verification, opens a PR - You review on GitHub like any other PR

You bring your own API keys (OpenAI, Anthropic, Gemini). We orchestrate but never store your code or prompts.

Free tier: 1 repo, 2 concurrent agents. Would love feedback from anyone who's tried Cursor/Copilot and wants something that works in parallel.

Comments URL: https://news.ycombinator.com/item?id=47111058

Points: 1

# Comments: 0

Hey HN,

I'm shipping my first open-source project and I'm pretty nervous about it. Aethene is an AI memory API – it gives your AI apps persistent memory. Store conversations, extract facts automatically, search semantically, handle contradictions gracefully. It works well thank most of the memory projects available on the market currently.

Why I built this: I was building AI agents and kept running into the same problem – they forget everything. Every conversation starts from zero. I wanted something that could: - Auto-extract facts from conversations (not just store raw text) - Handle "user moved from SF to NYC" without keeping both as true - Search by meaning, not just keywords - Version everything (who said what, when)

Tech stack: - TypeScript + Hono (fast, edge-ready) - Convex (real-time DB + vector search) - Gemini (embeddings + extraction) What it does: # Store memory curl -X POST /v1/content -d '{"content": "User loves hiking, lives in SF"}' # Recall naturally curl -X POST /v1/recall -d '{"query": "outdoor hobbies"}' # Returns: "User loves hiking" with assembled context It handles the boring stuff – chunking, embeddings, deduplication, contradiction detection, versioning – so you can focus on your actual product. Links: - GitHub: https://github.com/akhilponnada/aethene - API Docs: OpenAPI spec in repo This is my first time launching anything publicly. Would love feedback – what's missing? What would make you actually use this? Roast my code if you want, I can take it.

Thanks for reading.

Comments URL: https://news.ycombinator.com/item?id=47111048

Points: 1

# Comments: 0

I've been using OpenClaw (AI assistant framework) for a while and loved it, so I wanted to roll it out to my whole team. The problem: getting non-engineers set up is painful, and sharing API keys across a team is a security headache.

So I built ClawHuddle — a self-hosted platform that lets you provision and manage OpenClaw instances for your team from a single dashboard.

What it does:

- One-click provisioning — Admin invites a user, they get their own isolated OpenClaw instance (each runs in its own Docker container)

- Managed skills — Admin curates and pre-installs skills for the team, so everyone gets vetted tooling out of the box

- Centralized API keys — No need for every team member to have their own keys. Also supports Claude Code CLI token login

- Channel setup — Quickly connect messaging channels (Telegram supported now; Discord and LINE planned)

Each instance is fully isolated — conversations, files, and configs never leak between users.

Stack: Next.js 16, Fastify, SQLite, Docker, Traefik. Monorepo with Turborepo.

There's a live demo at https://clawhuddle.com/ (may have some rough edges, actively fixing bugs).

Open source: https://github.com/allen-hsu/clawhuddle

Hope this is useful if you're trying to bring OpenClaw to a team setting.

Comments URL: https://news.ycombinator.com/item?id=47111047

Points: 1

# Comments: 0

Your AI agent is burning 6x more tokens than it needs to just to browse the web. We built OpenBrowser MCP to fix that. Most browser MCPs give the LLM dozens of tools: click, scroll, type, extract, navigate. Each call dumps the entire page accessibility tree into the context window. One Wikipedia page? 124K+ tokens. Every. Single. Call. OpenBrowser works differently. It exposes one tool. Your agent writes Python code, and OpenBrowser executes it in a persistent runtime with full browser access. The agent controls what comes back. No bloated page dumps. No wasted tokens. Just the data your agent actually asked for. The result? We benchmarked it against Playwright MCP (Microsoft) and Chrome DevTools MCP (Google) across 6 real-world tasks: - 3.2x fewer tokens than Playwright MCP - 6x fewer tokens than Chrome DevTools MCP - 144x smaller response payloads - 100% task success rate across all benchmarks One tool. Full browser control. A fraction of the cost. It works with any MCP-compatible client: - Cursor - VS Code - Claude Code (marketplace plugin with MCP + Skills) - Codex and OpenCode (community plugins) - n8n, Cline, Roo Code, and more Install the plugins here: https://github.com/billy-enrizky/openbrowser-ai/tree/main/pl... It connects to any LLM provider: Claude, GPT 5.2, Gemini, DeepSeek, Groq, Ollama, and more. Fully open source under MIT license. OpenBrowser MCP is the foundation for something bigger. We are building a cloud-hosted, general-purpose agentic platform where any AI agent can browse, interact with, and extract data from the web without managing infrastructure. The full platform is coming soon. Join the waitlist at openbrowser.me to get free early access. See the Demo: https://youtu.be/ov1rSYd42hE?si=pB6QgtQfm-CX1CEa See the full benchmark methodology: https://docs.openbrowser.me/comparison See the benchmark code: https://github.com/billy-enrizky/openbrowser-ai/tree/main/be... Browse the source: https://github.com/billy-enrizky/openbrowser-ai LinkedIn Post: https://www.linkedin.com/posts/enrizky-brillian_opensource-a... #OpenSource #AI #MCP #BrowserAutomation #AIAgents #DevTools #LLM #GeneralPurposeAI #AgenticAI

Comments URL: https://news.ycombinator.com/item?id=47111045

Points: 1

# Comments: 0

Article URL: https://rename.world/

Comments URL: https://news.ycombinator.com/item?id=47111042

Points: 5

# Comments: 1