🛠️ All DevTools

[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] IPED Digital Forensic Tool. It is an open source software that can be used to process and analyze digital evidence, often seized at crime scenes by law enforcement or in a corporate investigation by private examiners.

[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 :)

[Testing] Test your prompts, agents, and RAGs. AI Red teaming, pentesting, and vulnerability scanning for LLMs. Compare performance of GPT, Claude, Gemini, Llama, and more. Simple declarative configs with command line and CI/CD integration.

[Other] OpenTelemetry for Rust Developers

[Other] Removing recursion via explicit callstack simulation

[Other] Claude Code, Claude Cowork and Codex #5

[Other] Show HN: I Was Here – Draw on street view, others can find your drawings Hey HN, I made a site where you can draw on street-level panoramas. Your drawings persist and other people can see them in real time.<p>Strokes get projected onto the 3D panorama so they wrap around buildings and follow the geometry, not just a flat overlay. Uses WebGL2 for rendering, Mapillary for the street imagery.<p>The idea is for it to become a global canvas, anyone can leave a mark anywhere and others stumble onto it.

[Other] The Cost of 'Lightweight' Frameworks: From Tauri to Native Rust

[Other] Oracle is building yesterday's data centers with tomorrow's debt

[Other] Code Review for Claude Code

[Other] Show HN: The Mog Programming Language Hi, Ted here, creator of Mog.<p>- Mog is a statically typed, compiled, embedded language (think statically typed Lua) designed to be written by LLMs -- the full spec fits in 3,200 tokens. - An AI agent writes a Mog program, compiles it, and dynamically loads it as a plugin, script, or hook. - The host controls exactly which functions a Mog program can call (capability-based permissions), so permissions propagate from agent to agent-written code. - Compiled to native code for low-latency plugin execution -- no interpreter overhead, no JIT, no process startup cost. - The compiler is written in safe Rust so the entire toolchain can be audited for security. Even without a full security audit, Mog is already useful for agents extending themselves with their own code. - MIT licensed, contributions welcome.<p>Motivations for Mog:<p>1. Syntax Only an AI Could Love: Mog is written for AIs to write, so the spec fits easily in context (~3200 tokens), and it&#x27;s intended to minimize foot-guns to lower the error rate when generating Mog code. This is why Mog has no operator precedence: non-associative operations have to use parentheses, e.g. (a + b) * c. It&#x27;s also why there&#x27;s no implicit type coercion, which I&#x27;ve found over the decades to be an annoying source of runtime bugs. There&#x27;s also less support in Mog for generics, and there&#x27;s absolutely no support for metaprogramming, macros, or syntactic abstraction.<p>When asking people to write code in a language, these restrictions could be onerous. But LLMs don&#x27;t care, and the less expressivity you trust them with, the better.<p>2. Capabilities-Based Permissionsl: There&#x27;s a paradox with existing security models for AI agents. If you give an agent like OpenClaw unfettered access to your data, that&#x27;s insecure and you&#x27;ll get pwned. But if you sandbox it, it can&#x27;t do most of what you want. Worse, if you run scripts the agent wrote, those scripts don&#x27;t inherit the permissions that constrain the agent&#x27;s own bash tool calls, which leads to pwnage and other chaos. And that&#x27;s not even assuming you run one of the many OpenClaw plugins with malware.<p>Mog tries to solve this by taking inspiration from embedded languages. It compiles all the way to machine code, ahead of time, but the compiler doesn&#x27;t output any dangerous code (at least it shouldn&#x27;t -- Mog is quite new, so that could still be buggy). This allows a host program, such as an AI agent, to generate Mog source code, compile it, and load it into itself using dlopen(), while maintaining security guarantees.<p>The main trick is that a Mog program on its own can&#x27;t do much. It has no direct access to syscalls, libc, or memory. It can basically call functions, do heap allocations (but only within the arena the host gives it), and return something. If the host wants the Mog program to be able to do I&#x2F;O, it has to supply the functions that the Mog program will call. A core invariant is that a Mog program should never be able to crash the host program, corrupt its state, or consume more resources than the host allows.<p>This allows the host to inspect the arguments to any potentially dangerous operation that the Mog program attempts, since it&#x27;s code that runs in the host. For example, a host agent could give a Mog program a function to run a bash command, then enforce its own session-level permissions on that command, even though the command was dynamically generated by a plugin that was written without prior knowledge of those permission settings.<p>(There are a couple other tricks that PL people might find interesting. One is that the host can limit the execution time of the guest program. It does this using cooperative interrupt polling, i.e. the compiler inserts runtime checks that check if the host has asked the guest to stop. This causes a roughly 10% drop in performance on extremely tight loops, which are the worst case. It could almost certainly be optimized.)<p>3. Self Modification Without Restart: When I try to modify my OpenClaw from my phone, I have to restart the whole agent. Mog fixes this: an agent can compile and run new plugins without interrupting a session, which makes it dynamically responsive to user feedback (e.g., you tell it to always ask you before deleting a file and without any interruption it compiles and loads the code to... actually do that).<p>Async support is built into the language, by adapting LLVM&#x27;s coroutine lowering to our Rust port of the QBE compiler, which is what Mog uses for compilation. The Mog host library can be slotted into an async event loop (tested with Bun), so Mog async calls get scheduled seamlessly by the agent&#x27;s event loop. Another trick is that the Mog program uses a stack inside the memory arena that the host provides for it to run in, rather than the system stack. The system tracks a guard page between the stack and heap. This design prevents stack overflow without runtime overhead.<p>Lots of work still needs to be done to make Mog a &quot;batteries-included&quot; experience like Python. Most of that work involves fleshing out a standard library to include things like JSON, CSV, Sqlite, and HTTP. One high-impact addition would be an `llm` library that allows the guest to make LLM calls through the agent, which should support multiple models and token budgeting, so the host could prevent the plugin from burning too many tokens.<p>I suspect we&#x27;ll also want to do more work to make the program lifecycle operations more ergonomic. And finally, there should be a more fully featured library for integrating a Mog host into an AI agent like OpenClaw or OpenAI&#x27;s Codex CLI.