🛠️ Hacker News Tools

[Other] Searching for the Agentic IDE <a href="https:&#x2F;&#x2F;xcancel.com&#x2F;karpathy&#x2F;status&#x2F;2031616709560610993" rel="nofollow">https:&#x2F;&#x2F;xcancel.com&#x2F;karpathy&#x2F;status&#x2F;2031616709560610993</a>

[DevOps] Show HN: Ink – Deploy full-stack apps from AI agents via MCP or Skills Hi HN, I built Ink, a full stack deployment platform where the primary users are AI agents, not humans.<p>We all know AI can write code, but deploying them still requires a human to wire it up: hosting, databases, DNS, and secrets. Ink gives agents those tools directly.<p>The agent calls &quot;deploy&quot; and the platform auto-detects the framework, builds it, deploys it, and returns a live URL at *.ml.ink. Here&#x27;s a demo with Claude Code: <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=F6ZM_RrIaC0" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=F6ZM_RrIaC0</a>.<p>What Ink does that I haven&#x27;t seen elsewhere:<p>- One agent skill for compute + databases + DNS + secrets + domains + usage + metrics + logs + scaling. The agent doesn&#x27;t juggle separate providers — one account, one auth, one set of tools.<p>- DNS zone delegation. Delegate a zone once (e.g. dev.acme.com) and agents create any subdomain instantly — no manual adding DNS records each time, no propagation wait.<p>- Multiple agents and humans share one workspace and collaborate on projects. I envision a future where many agents collaborate together. I&#x27;m working on a cool demo to share.<p>- Built-in git hosting. Agents push code and deploy without the human setting up GitHub first. No external account needed. (Of course if you&#x27;re a developer you can store code on GitHub — that&#x27;s the recommended pattern.)<p>You also have what you&#x27;d expect: - UI with service observability designed for humans (logs, metrics, DNS). - GitHub integration — push triggers auto-redeploy. - Per-minute billing for CPU, memory, and egress. No per-seat, no per-agent. - Error responses designed for LLMs. Structured reason codes with suggested next actions, not raw stack traces. When a deploy fails the agent reads the log, fixes it, and redeploys autonomously.<p>Try: <a href="https:&#x2F;&#x2F;ml.ink" rel="nofollow">https:&#x2F;&#x2F;ml.ink</a> Free $2 trial credits, no credit card. In case you want to try further here&#x27;s 20% code &quot;GOODFORTUNE&quot;.

[Other] Show HN: OpenUI – A code-like rendering spec for Generative UI Thesys just open-sourced their generative UI rendering engine. Interesting timing given where Google a2ui and Vercel&#x27;s json-render are headed. The difference worth noting: a2ui and json-render both treat JSONL as the contract between the LLM and the renderer. Thesys is betting that&#x27;s the wrong primitive. Their engine uses a code-like syntax (OpenUI Lang) instead — LLM writes it, renderer executes it. The argument is that LLMs are fundamentally better at generating code than generating structured data, so you get cleaner output and ~67% fewer tokens. The broader vision seems to be a model-agnostic, design-system-agnostic layer that sits between any LLM and your actual UI components. You bring your own components and design tokens, the engine handles translating LLM output into rendered interfaces — charts, forms, tables, cards. Generative UI as a category is still figuring out what the right abstraction is. This is a concrete stake in the ground against JSON-as-spec.

[Other] Show HN: Open-source browser for AI agents Hi HN, I forked chromium and built agent-browser-protocol (ABP) after noticing that most browser-agent failures aren’t really about the model misunderstanding the page. Instead, the problem is that the model is reasoning from a stale state.<p>ABP is designed to keep the acting agent synchronized with the browser at every step. After each action (click, type, etc), it freezes JavaScript execution and rendering, then captures the resulting state. It also compiles the notable events that occurred during that action loop, such as navigation, file pickers, permission prompts, alerts, and downloads, and sends that along with a screenshot of the frozen page state back to the agent.<p>The result is that browser interaction starts to feel more like a multimodal chat loop. The agent takes an action, gets back a fresh visual state and a structured summary of what happened, then decides what to do next from there. That fits much better with how LLMs already work.<p>A few common browser-use failures ABP helps eliminate: * A modal appears after the last Playwright screenshot and blocks the input the agent was about to use * Dynamic filters cause the page to reflow between steps * An autocomplete dropdown opens and covers the element the agent intended to click * alert() &#x2F; confirm() interrupts the flow * Downloads are triggered, but the agent has no reliable way to know when they’ve completed<p>As proof, ABP with opus 4.6 as the driver scores 90.5% on the Online Mind2Web benchmark. I think modern LLMs already understand websites, they just need a better tool to interact with them. Happy to answer questions about the architecture, forking chrome or anything else in the comments below.<p>Try it out: `claude mcp add browser -- npx -y agent-browser-protocol --mcp` (Codex&#x2F;OpenCode instructions in the docs)<p>Demo video: <a href="https:&#x2F;&#x2F;www.loom.com&#x2F;share&#x2F;387f6349196f417d8b4b16a5452c3369" rel="nofollow">https:&#x2F;&#x2F;www.loom.com&#x2F;share&#x2F;387f6349196f417d8b4b16a5452c3369</a>

[DevOps] Show HN: I built an ISP infrastructure emulator from scratch with a custom vBNG Demo: <a href="https:&#x2F;&#x2F;aether.saphal.me" rel="nofollow">https:&#x2F;&#x2F;aether.saphal.me</a> GitHub: <a href="https:&#x2F;&#x2F;github.com&#x2F;saphalpdyl&#x2F;Aether" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;saphalpdyl&#x2F;Aether</a><p>Aether is a multi-BNG (Broadband Network Gateway) ISP infrastructure lab built almost from scratch that emulates IPoE IPv4 subscriber management end-to-end. It supports IPoE&#x2F;Ipv4 networks and runs a python-based vBNG with RADIUS AAA, per-subscriber traffic shaping, and traffic simulation emulated on Containerlab. It is also my first personal networking project, built roughly over a month.<p>Motivations behind the project<p>I&#x27;m a CS sophomore. About three years ago, I was assigned, as an intern, to build a OSS&#x2F;BSS platform for a regional ISP by myself without mentoring. Referencing demo.splynx.com , I developed most of the BSS side ( bookkeeping, accounting, inventory management ), but, in terms of networking, I managed to install and setup RADIUS and that was about it. I didn&#x27;t have anyone to mentor me or ask questions to, so I had given up then.<p>Three years later, I decided to try cracking it again. This project is meant to serve as a learning reference for anyone who&#x27;s been in that same position i.e staring at closed-source vendor stacks without proper guidance. This is absolutely not production-grade, but I hope it gives someone a place to start.<p>Architecture overview<p>The core component, the BNG, runs on an event-driven architecture where state changes are passed around as messages to avoid handling mutexes and locks. The session manager is the sole owner of the session state. To keep it clean and predictable, the direBNG never accepts external inputctly. The one exception is the Go RADIUS CoA daemon, which passes CoA messages in via IPC sockets. Everything the BNG produces(events, session snapshots) gets pushed to Redis Streams, where the bng-ingestor picks them up, processes them, and persists them.<p>Simulation and meta-configs<p>I am generating traffic through a simulator node that mounts the host&#x27;s docker socket and runs docker exec commands on selected hosts. The topology.yaml used by Containerlab to define the network topology grows bigger as more BNG&#x27;s and access nodes are added. So aether.config.yaml, a simpler configuration, is consumed by the configuration pipeline to generate the topology.yaml and other files (nginx.conf, kea-dhcp.conf, RADIUS clients.conf etc.)<p>Known Limitations<p>- Multiple veth hops through the emulated topology add significant overhead. Profiling with iperf3 (-P 10 -t 10, 9500 MTU, 24 vCPUs) shows BNG→upstream at ~24 Gbit&#x2F;s, but host→BNG→upstream drops to ~3.5 Gbit&#x2F;s. The 9500 MTU also isn&#x27;t representative of real ISP deployments. This gets worse when the actual network is reintroduced capping my throughput to 1.6 Gbits&#x2F;sec in local. - The circuit ID format (1&#x2F;0&#x2F;X) is non-standard. I simplified it for clarity. - No iBGP or VLAN support. - No Ipv6 support. I wanted to target IPv4 networks from the start to avoid getting too much breadth without a lot of depth.<p>Nearly everything I know about networking (except some sections from AWS) I learned building this. A lot was figured out on the fly, so engineers will likely spot questionable decisions in the codebase. I&#x27;d genuinely appreciate that feedback.<p>Questions<p>- Currently, the circuit where the user connects is arbitrarily decided by the demo user. In a real system with thousands of circuits, it&#x27;d be very difficult to properly assess which circuit the customer might connect to. When adding a new customer to a service, how does the operator decide, based on customer&#x27;s location, which circuit to provide the service to ?

[Other] BitNet: Inference framework for 1-bit LLMs Paper: <a href="https:&#x2F;&#x2F;arxiv.org&#x2F;pdf&#x2F;2310.11453" rel="nofollow">https:&#x2F;&#x2F;arxiv.org&#x2F;pdf&#x2F;2310.11453</a>

[Other] Show HN: Modulus – Cross-repository knowledge orchestration for coding agents Hello HN, we&#x27;re Jeet and Husain from Modulus (<a href="https:&#x2F;&#x2F;modulus.so" rel="nofollow">https:&#x2F;&#x2F;modulus.so</a>) - a desktop app that lets you run multiple coding agents with shared project memory. We built it to solve two problems we kept running into:<p>- Cross-repo context is broken. When working across multiple repositories, agents don&#x27;t understand dependencies between them. Even if we open two repos in separate Cursor windows, we still have to manually explain the backend API schema while making changes in the frontend repo.<p>- Agents lose context. Switching between coding agents often means losing context and repeating the same instructions again.<p>Modulus shares memory across agents and repositories so they can understand your entire system.<p>It&#x27;s an alternative to tools like Conductor for orchestrating AI coding agents to build product, but we focused specifically on multi-repo workflows (e.g., backend repo + client repo + shared library repo + AI agents repo). We built our own Memory and Context Engine from the ground up specifically for coding agents.<p>Why build another agent orchestration tool? It came from our own problem. While working on our last startup, Husain and I were working across two different repositories. Working across repos meant manually pasting API schemas between Cursor windows — telling the frontend agent what the backend API looked like again and again. So we built a small context engine to share knowledge across repos and hooked it up to Cursor via MCP. This later became Modulus.<p>Soon, Modulus will allow teams to share knowledge with others to improve their workflows with AI coding agents - enabling team collaboration in the era of AI coding. Our API will allow developers to switch between coding agents or IDEs without losing any context.<p>If you wanna see a quick demo before trying out, here is our launch post - <a href="https:&#x2F;&#x2F;x.com&#x2F;subhajitsh&#x2F;status&#x2F;2024202076293841208" rel="nofollow">https:&#x2F;&#x2F;x.com&#x2F;subhajitsh&#x2F;status&#x2F;2024202076293841208</a><p>We&#x27;d greatly appreciate any feedback you have and hope you get the chance to try out Modulus.

[API/SDK] AEP (API Design Standard and Tooling Ecosystem)

[Other] Launch HN: RunAnywhere (YC W26) – Faster AI Inference on Apple Silicon Hi HN, we&#x27;re Sanchit and Shubham (YC W26). We built a fast inference engine for Apple Silicon. LLMs, speech-to-text, text-to-speech – MetalRT beats llama.cpp, Apple&#x27;s MLX, Ollama, and sherpa-onnx on every modality we tested. Custom Metal shaders, no framework overhead.<p>Also, we&#x27;ve open-sourced RCLI, the fastest end-to-end voice AI pipeline on Apple Silicon. Mic to spoken response, entirely on-device. No cloud, no API keys.<p>To get started:<p><pre><code> brew tap RunanywhereAI&#x2F;rcli https:&#x2F;&#x2F;github.com&#x2F;RunanywhereAI&#x2F;RCLI.git brew install rcli rcli setup # downloads ~1 GB of models rcli # interactive mode with push-to-talk </code></pre> Or:<p><pre><code> curl -fsSL https:&#x2F;&#x2F;raw.githubusercontent.com&#x2F;RunanywhereAI&#x2F;RCLI&#x2F;main&#x2F;install.sh | bash </code></pre> The numbers (M4 Max, 64 GB, reproducible via `rcli bench`):<p>LLM decode – 1.67x faster than llama.cpp, 1.19x faster than Apple MLX (same model files): - Qwen3-0.6B: 658 tok&#x2F;s (vs mlx-lm 552, llama.cpp 295) - Qwen3-4B: 186 tok&#x2F;s (vs mlx-lm 170, llama.cpp 87) - LFM2.5-1.2B: 570 tok&#x2F;s (vs mlx-lm 509, llama.cpp 372) - Time-to-first-token: 6.6 ms<p>STT – 70 seconds of audio transcribed in *101 ms*. That&#x27;s 714x real-time. 4.6x faster than mlx-whisper.<p>TTS – 178 ms synthesis. 2.8x faster than mlx-audio and sherpa-onnx.<p>We built this because demoing on-device AI is easy but shipping it is brutal. Voice is the hardest test: you&#x27;re chaining STT, LLM, and TTS sequentially, and if any stage is slow, the user feels it. Most teams fall back to cloud APIs not because local models are bad, but because local inference infrastructure is.<p>The thing that&#x27;s hard to solve is latency compounding. In a voice pipeline, you&#x27;re stacking three models in sequence. If each adds 200ms, you&#x27;re at 600ms before the user hears a word, and that feels broken. You can&#x27;t optimize one stage and call it done. Every stage needs to be fast, on one device, with no network round-trip to hide behind.<p>We went straight to Metal. Custom GPU compute shaders, all memory pre-allocated at init (zero allocations during inference), and one unified engine for all three modalities instead of stitching separate runtimes together.<p>MetalRT is the first engine to handle all three modalities natively on Apple Silicon. Full methodology:<p>LLM benchmarks: <a href="https:&#x2F;&#x2F;www.runanywhere.ai&#x2F;blog&#x2F;metalrt-fastest-llm-decode-engine-apple-silicon">https:&#x2F;&#x2F;www.runanywhere.ai&#x2F;blog&#x2F;metalrt-fastest-llm-decode-e...</a><p>Speech benchmarks: <a href="https:&#x2F;&#x2F;www.runanywhere.ai&#x2F;blog&#x2F;metalrt-speech-fastest-stt-tts-apple-silicon">https:&#x2F;&#x2F;www.runanywhere.ai&#x2F;blog&#x2F;metalrt-speech-fastest-stt-t...</a><p>How: Most inference engines add layers between you and the GPU: graph schedulers, runtime dispatchers, memory managers. MetalRT skips all of it. Custom Metal compute shaders for quantized matmul, attention, and activation - compiled ahead of time, dispatched directly.<p>Voice Pipeline optimizations details: <a href="https:&#x2F;&#x2F;www.runanywhere.ai&#x2F;blog&#x2F;fastvoice-on-device-voice-ai-pipeline-apple-silicon">https:&#x2F;&#x2F;www.runanywhere.ai&#x2F;blog&#x2F;fastvoice-on-device-voice-ai...</a> RAG optimizations: <a href="https:&#x2F;&#x2F;www.runanywhere.ai&#x2F;blog&#x2F;fastvoice-rag-on-device-retrieval-augmented-voice-ai">https:&#x2F;&#x2F;www.runanywhere.ai&#x2F;blog&#x2F;fastvoice-rag-on-device-retr...</a><p>RCLI is the open-source voice pipeline (MIT) built on MetalRT: three concurrent threads with lock-free ring buffers, double-buffered TTS, 38 macOS actions by voice, local RAG (~4 ms over 5K+ chunks), 20 hot-swappable models, and a full-screen TUI with per-op latency readouts. Falls back to llama.cpp when MetalRT isn&#x27;t installed.<p>Source: <a href="https:&#x2F;&#x2F;github.com&#x2F;RunanywhereAI&#x2F;RCLI" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;RunanywhereAI&#x2F;RCLI</a> (MIT)<p>Demo: <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=eTYwkgNoaKg" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=eTYwkgNoaKg</a><p>What would you build if on-device AI were genuinely as fast as cloud?

[Other] Show HN: Agentic Data Analysis with Claude Code Hey HN, as a former data analyst, I’ve been tooling around trying to get agents to do my old job. The result is this system that gets you maybe 80% of the way there. I think this is a good data point for what the current frontier models are capable of and where they are still lacking (in this case — hypothesis generation and general data intuition).<p>Some initial learnings: - Generating web app-based reports goes much better if there are explicit templates&#x2F;pre-defined components for the model to use. - Claude can “heal” broken charts if you give it access to chart images and run a separate QA loop.<p>Would either feedback from the community or to hear from others that have tried similar things!

[Other] I built a programming language using Claude Code

[Other] Show HN: A modern React onboarding tour library react-tourlight is the modern React tour library. Zero dependencies, WCAG 2.1 AA accessible, under 5 kB gzipped. The one that works with React 19.

[Other] Show HN: Ash, an Agent Sandbox for Mac Ash is a macOS sandbox that restricts AI coding agents. It limits access to files, networks, processes, IO devices, and environment variables. You can use Ash with any CLI coding agent by wrapping it in a single command: `ash run -- &lt;agent&gt;`. I typically use it with Claude to stay safe while avoiding repetitive prompts: `ash run -- claude --dangerously-skip-permissions`.<p>Ash restricts resources via the Endpoint Security and Network Extension frameworks. These frameworks are significantly more powerful than the sandbox-exec tool.<p>Each session is driven by a policy file. Any out-of-policy action is denied by default. You can audit denials in the GUI app, which lets you view out-of-policy actions and retroactively add them to your policy file.<p>Ash also comes with tools for building policies. You can use an &quot;observation session&quot; to watch the typical behavior of a coding agent and capture that behavior in a policy file for future sandbox sessions. Linting, formatting, and rule merging are all built into the Ash CLI to keep your policy files concise and maintainable.<p>Download Ash at <a href="https:&#x2F;&#x2F;ashell.dev" rel="nofollow">https:&#x2F;&#x2F;ashell.dev</a>

[Other] Rebasing in Magit

[Other] Show HN: How I topped the HuggingFace open LLM leaderboard on two gaming GPUs I found that duplicating a specific block of 7 middle layers in Qwen2-72B, without modifying any weights, improved performance across all Open LLM Leaderboard benchmarks and took #1. As of 2026, the top 4 models on that leaderboard are still descendants.<p>The weird finding: single-layer duplication does nothing. Too few layers, nothing. Too many, it gets worse. Only circuit-sized blocks of ~7 layers work. This suggests pretraining carves out discrete functional circuits in the layer stack that only work when preserved whole.<p>The whole thing was developed on 2x RTX 4090s in my basement. I&#x27;m now running current models (GLM-4.7, Qwen3.5, MiniMax M2.5) on a dual GH200 rig (see my other post). Code and new models coming soon.<p>Happy to answer questions.

[CLI Tool] Show HN: Smux – Terminal Multiplexer built for AI agents

[Other] Show HN: DD Photos – open-source photo album site generator (Go and SvelteKit) I was frustrated with photo sharing sites. Apple&#x27;s iCloud shared albums take 20+ seconds to load, and everything else comes with ads, cumbersome UIs, or social media distractions. I just want to share photos with friends and family: fast, mobile-friendly, distraction-free.<p>So I built DD Photos. You export photos from whatever you already use (Lightroom, Apple Photos, etc.) into folders, run `photogen` (a Go CLI) to resize them to WebP and generate JSON indexes, then deploy the SvelteKit static site anywhere that serves files. Apache, S3, whatever. No server-side code, no database.<p>Built over several weeks with heavy use of Claude Code, which I found genuinely useful for this kind of full-stack project spanning Go, SvelteKit&#x2F;TypeScript, Apache config, Docker, and Playwright tests. Happy to discuss that experience too.<p>Live example: <a href="https:&#x2F;&#x2F;photos.donohoe.info" rel="nofollow">https:&#x2F;&#x2F;photos.donohoe.info</a> Repo: <a href="https:&#x2F;&#x2F;github.com&#x2F;dougdonohoe&#x2F;ddphotos" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;dougdonohoe&#x2F;ddphotos</a>

[Other] Show HN: Local-first firmware analyzer using WebAssembly Hi HN,<p>I just wanted to share what I have been working on for the past few months: A firmware analyzer for embedded Linux systems that helps uncovering security issues running entirely in the browser.<p>This is a very early Alpha. It is going to be rough around the edges. But I think it provides quite a lot of value already.<p>So please go ahead and drop a firmware (only .tar rootfs archives for now) and try to break it :)

[Other] OpenTelemetry for Rust Developers

[Other] Removing recursion via explicit callstack simulation