🛠️ Hacker News Tools

Show HN: Frugon – Find which LLM calls a cheaper model could handle (local, MIT) I started leaning in on AI heavily this year, as I wanted to get more done autonomously, but then my token usage climbed dramatically to the point where my weekly quota would run out before the end of the week, sometimes a couple of days into the week.<p>I realised I had to do something about it else I&#x27;d have to double my spend. So I decided to start tracking my cost per task type. This revealed that a lot of my spend went to searches&#x2F;scans or simple things like scouting tasks.<p>I then decided to turn this into a simple CLI tool that can be used to read your OpenAI-style logs locally, and analyze the cost and compare this spend to other models, then show you how much you could potentially save by switching those calls to a cheaper model.<p>When you run analyze you get an offline estimate priced against LiteLLM and gated by LMArena tiers. The general savings bands come from the research published by RouteLLM; but you can confirm this yourself using 2 commands --measure (shows the prompt-response output side by side) and --judge (a model chosen to do the comparisons). These send a sample of the prompts from the logs to the candidate models - either the default choice or set by you. This call goes directly to the model provider (never through me) as any normal LLM call would, and the response is shown and judged to either be better or worse or a tie.<p>It&#x27;s deliberately small, because I tend to over complicate&#x2F;think things sometimes: analyze + capture + a few commands, doing three jobs. Cost, quality visibility, routing recommendation.<p>Nothing is hosted. capture is an optional local proxy on your own machine, and there&#x27;s no endpoint in the path of your data. You can confirm this by checking the source.<p>I included a demo so you can check out the output. It has a synthetic 56k call log (a month&#x27;s worth) showing how costs can drop from $549.46 to $343.91 a month. A 37.4% saving.<p>Try it:<p><pre><code> uvx frugon analyze --demo </code></pre> or<p><pre><code> uv tool install frugon </code></pre> Then point it at your own logs.<p>All feedback is welcome, especially any on the routing&#x2F;quality logic, or anything else, good or bad.

Show HN: Fast, native Mac file manager (filters, fuzzy find, 9 MB, no Electron) My Downloads folder had been left unkept for a really long time and cleaning it up using Finder was quite cumbersome. So I started creating a simple app to help me filter out and delete or move the files in the folder.<p>It started out very basic and the filtering options genuinely helped me clean out the Downloads folder, then as I thought of more features I would like to see in a file manager I started to add them. Some of the features are:<p>- Fuzzy go to folder&#x2F;file where you only need to write a few letters of a full path, get suggestions and can jump to the correct path instantly - Hover over a file to preview an image or PDF without opening it - Dual-pane view and tabs that remember selected filters - Command palette to find actions fast - Batch rename (with regex support and presets for common operations), image conversion (HEIC&#x2F;WebP&#x2F;AVIF to JPG&#x2F;PNG), zip creation - Bookmarks, Quick Look, single-click open, keyboard control<p>And many other small quality of life features.<p>I used Claude Code for making the app and the tech stack is .NET&#x2F;C# with AppKit. The app is compiled to Native AOT so the total app size is only around 9 MB. Since file operations are very important to get right (I don’t want to lose any important files) I put a lot of time into hardening file move&#x2F;copy&#x2F;delete operations. Copies are written to a temp file and atomically renamed into place. I also made a dedicated audit of move&#x2F;copy&#x2F;delete and verified each operation by hand.<p>Another thing that’s important to me is apps that are privacy first so the only network request the app makes is to check if there’s a new version (this is not done at all for the trial version, the trial expiry date is enforced locally by just checking the date and trusting the user) and the only information that is sent is the current app version.<p>The app is Apple Silicon, macOS 12+ and has a 30-day free trial with a one-time launch price of $19.99 (no subscription).

The Making of Claude Code

Inference Optimization for MiMo v2.5: Pushing Hybrid SWA Efficiency to the Limit

Cache-Conscious Data Layout in Rust: Field Zoning, False Sharing, 128-Byte Rule

Apache Shiro security framework releases 3.0.0

Show HN: Capsule – E2E encrypted file sharing Capsule started as a basic HTTP server to teach myself Rust.<p>Then I kept going, and it turned into the first full-stack thing I&#x27;ve fully shipped: a Rust&#x2F;Axum server, a Rust CLI, an Android app (my first, same story as the Rust one), and a fully featured web UI. It quickly grew as I realized I could solve my file transfer problem by turning a basic webserver into an actual daily use tool.<p>Entire project is open source, code can be found at: <a href="https:&#x2F;&#x2F;github.com&#x2F;withcapsule&#x2F;" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;withcapsule&#x2F;</a><p>Would love to hear anyone&#x27;s thoughts on this.

Show HN: InstantVideos.org – short documentaries in ~30 seconds Hiya! So I&#x27;ve been playing around with having Claude make videos for a bit now even had some success posting the results to TikTok (and setup a whole pipeline so Claude can generate and post autonomously). With the release of Nano Banana 2 Lite, I was curious show fast I could make the generation, so last night I gave it a whirl and got down to around 30s for short-form video.<p>It uses GLM-5.2 fast via Fireworks to generate the scripts and image prompts and, like I said, Nano Banana 2 Lite for the images, gpt-4o-mini-tts for the narration, and ffmpeg to string it all together and add the Ken Burns zoom effect (which still has a shake I haven&#x27;t been able to get rid of). The video compilation proved to be the blocker once the rest was in place, but I was able to speed that up by putting it on a 64 vCPU EC2.<p>The cost might be the most interesting aspect as the short form videos tend to be about 25 cents. Almost 90% of that is the images, which are 3.336 cents a piece. Of course, running the big 64 core EC2 to allow for the creation isn&#x27;t cheap.<p>It seems like on-demand AI video is coming, and I thought this was an interesting demo of how close it might be in at least one narrow video domain.

AI: The ROI Runway Could Be Long Outside the Tech Sector

Learning to code is still worthwhile

Poison, redzones and shadows: inside KASAN

Python 3.14 compiled to metal – no interpreter

Show HN: orzma – a terminal emulator that renders webviews inside the terminal I made this because I thought it would be useful to be able to use a webview in the terminal. I also provide an SDK called ratatui_orzma. For example, I think it could be used to render rich UI components such as charts with a webview as part of a TUI application, port web tools to ratatui, or embed games built with Wasm inside the terminal.

Show HN: An interactive, interlinked city exploration project (St. Louis) I moved to St. Louis a few years ago, and I started documenting the food I&#x27;ve tried, neighborhoods I&#x27;ve explored, and just the random stuff I&#x27;ve learned in a website.<p>It has a few features:<p>Food map[1]: a map where you can filter restaurants&#x2F;cafes by type of cuisine, neighborhood, or rating, and sort by rating or alphabetical.<p>Neighborhood map[2]: a map of all the neighborhoods and parks in the city, and some notable St. Louis county suburbs.<p>I tried to make it interlinked and fun to explore, so you can land on the page for Sultan (restaurant) which leads you to The Grove (neighborhood), which could lead you back to Grace Meat &amp; Three (nearby restaurant), or over to Midtown (nearby neighborhood), etc.<p>Built with Astro, hosted on Github[3], self-hosted map tiles.<p>[1]: <a href="https:&#x2F;&#x2F;stlouing.com&#x2F;food&#x2F;" rel="nofollow">https:&#x2F;&#x2F;stlouing.com&#x2F;food&#x2F;</a><p>[2]: <a href="https:&#x2F;&#x2F;stlouing.com&#x2F;neighborhoods&#x2F;" rel="nofollow">https:&#x2F;&#x2F;stlouing.com&#x2F;neighborhoods&#x2F;</a><p>[3]: <a href="https:&#x2F;&#x2F;github.com&#x2F;stlouing&#x2F;stlouing.com" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;stlouing&#x2F;stlouing.com</a>

Show HN: Focal Harvest – CLI research pipeline to search, scrape and notify

OfficeCLI: Office suite for AI agents to read and edit Microsoft Office files

Show HN: Pulpie – Models for Cleaning the Web Hey HN, I&#x27;m Shreyash, founder of Feyn. We built Pulpie, a family of Pareto optimal models for cleaning the web. Pulpie strips boilerplate (ads, footers, sidebars) from raw HTML and returns just the main content as HTML or Markdown.<p>We match SOTA extraction quality while being 20x cheaper. Cleaning 1 billion webpages costs $7,900 with Pulpie versus $159,000 with Dripper, the current leading extractor.<p>The gains come from architecture. Today&#x27;s leading extractors are decoders that generate output one token at a time. Each step reads the full model from memory to produce a single token. Conversely, Pulpie models are encoders. They run one forward pass over the full input HTML and label each block as boilerplate or content. As a result, Pulpie is compute-bound while decoders are memory-bound. Cheaper GPUs have relatively more compute than memory bandwidth. This makes Pulpie easy to run optimally.<p>Here&#x27;s Pulpie and Dripper cleaning the same pages side by side: <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=ibd-tIiQECo" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=ibd-tIiQECo</a>. You can try a side-by-side comparison yourself: <a href="https:&#x2F;&#x2F;huggingface.co&#x2F;spaces&#x2F;feyninc&#x2F;pulpie" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;spaces&#x2F;feyninc&#x2F;pulpie</a><p>Our motivation for Pulpie came from building a deep research harness. Every search API returns noisy content that contains ads, nav elements, and sidebars. In one instance, an ad for &quot;Gemini on Pixel&quot; slipped into our search results, got passed into LLM context, and ended up in the final answer served to the user. Pretty embarrassing moment for us but it helped us realize how bad data kills model intelligence. We built Pulpie to get clean data for cheap.<p>All models are open source on Hugging Face. You can read about our training process and how to use Pulpie here: <a href="https:&#x2F;&#x2F;usefeyn.com&#x2F;blog&#x2F;pulpie-pareto-optimal-models-for-cleaning-the-web&#x2F;#get-started">https:&#x2F;&#x2F;usefeyn.com&#x2F;blog&#x2F;pulpie-pareto-optimal-models-for-cl...</a><p>Happy to answer any questions!

CS2 Fog Of War: Server-sided anti-wallhack occlusion culling for CS2 servers

Show HN: Nectar, a Rust-like React that compiles to WebAssembly After a pretty rough gig experience, I wanted to remind myself that I was a vaguely capable developer. I had recently jumped on the Rust train and wanted a Rust equivalent of React that is more performant and easier to work with. I also wondered if I could effectively eject Javascript. This project came from that... Nectar is a programming language that compiles your entire app, logic, state, and rendering, to WebAssembly. JavaScript is reduced to a roughly 10 KB syscall layer that only bridges WASM to the DOM.<p>The origin: I set out to get React-style ergonomics with better performance, and about halfway in I realized I had hit the ceiling of what was possible going through the browser&#x27;s DOM (Blink&#x2F;C++) and a diffing step. So I stopped fighting it and built my own signal-driven rendering layer that compiles to WASM and updates exactly the DOM nodes that changed, in O(1), with no virtual DOM and no garbage collector.<p>What it is right now:<p>- One binary. The same compiler does formatting, linting, testing, a dev server, an LSP, and SSR (although SSR is underdeveloped). - Fine-grained signals. Each binding updates on its own, so there is no diffing pass. - A Rust-inspired borrow checker for compile-time safety. - Language-level keywords for the patterns you reach for constantly on the web: component, store, router, form, auth, payment. - SEO and Screen Readers. It compiles to two DOMs from the same source so the app stays visible to screen readers and crawlers.<p>Honest status, because this is early and it is mostly one person:<p>- The &quot;no JavaScript&quot; line has an asterisk. WASM cannot touch the DOM directly, so there is a small JS bridge (about 10 KB) that as of this moment I cannot overcome on major browsers. - The numbers on the site (a counter is ~10 KB versus ~387 KB for React plus JS) are real but small examples, not a full-app shootout. I would not read them as more than they are. They are intentionally poorly architected to measure throughput at massive scales. (contrived) - License is BSL 1.1, so read that before you build anything real on it.<p>Live demo, 10,000 products with a canvas rendering engine, plus a side by side against Svelte 5: <a href="https:&#x2F;&#x2F;buildnectar.com&#x2F;app&#x2F;canvas" rel="nofollow">https:&#x2F;&#x2F;buildnectar.com&#x2F;app&#x2F;canvas</a><p>Site: <a href="https:&#x2F;&#x2F;buildnectar.com" rel="nofollow">https:&#x2F;&#x2F;buildnectar.com</a> Code: <a href="https:&#x2F;&#x2F;github.com&#x2F;HibiscusConsulting&#x2F;nectar-lang" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;HibiscusConsulting&#x2F;nectar-lang</a><p>I would genuinely like it torn apart. If the core bet is wrong, crush my dreams but tell me how we can make it better, please. I still am not completely sure this is even a good idea. I honestly thought this would be a lot easier to build than it has been. If I had known at the start, I would have never built it, but here we are.

Show HN: Agent Draw: An agent draws while you talk, built on TLDraw We recently built 2draw, a Drawful-style game where players draw on a shared canvas and race to guess each other&#x27;s drawings, on tldraw, an infinite-canvas SDK for React.<p>We started wondering what it would take to put an agent in that loop, as an opponent or a rival guesser, and dug into how an agent could read and draw on a tldraw canvas. That research turned into:<p>Agent draw, a tool that lets an agent draw to the canvas for you while you present.<p>You can try it right now, or grab the source:<p>- Live demo: <a href="https:&#x2F;&#x2F;tldraw-agent-draw-demo.james-664.workers.dev" rel="nofollow">https:&#x2F;&#x2F;tldraw-agent-draw-demo.james-664.workers.dev</a><p>- Source: <a href="https:&#x2F;&#x2F;github.com&#x2F;ritza-co&#x2F;tldraw-agent-draw-demo" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;ritza-co&#x2F;tldraw-agent-draw-demo</a><p>Here you can see the agent assisting me in my demo presentation of a third grade chemistry class:<p>- <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=iIH2hJAxxm8" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=iIH2hJAxxm8</a>