2025年8月12日 の Show HN

Hey ya’ll, Ishaan and Kartik here. We're building Omnara (https://omnara.com/), an “agent command center” that lets you launch and control Claude Code from anywhere: terminal, web, or mobile — and easily switch between them. Check out a demo here: https://www.loom.com/share/03d30efcf8e44035af03cbfebf840c73.

Run 'pip install omnara && omnara', and you'll have a regular Claude Code session. But you can continue that same session from our web dashboard (https://omnara.com/) or mobile app (https://apps.apple.com/us/app/omnara-ai-command-center/id674...).

Before Omnara, we felt stuck watching Claude Code think and write code, waiting 5-10 minutes just to provide input when needed. Now with Omnara, I can start a Claude Code session and if I need to leave my laptop, I can respond from my phone anywhere. Some places I've coded from include my bed, on a walk, in an Uber, while doing laundry, and even on the toilet.

There are many new Claude Code wrappers (e.g., Crystal, Conductor), but none keep the native Claude Code terminal experience while allowing interaction outside the terminal, especially on mobile. On the other hand, tools like Vibetunnel or Termius replicate the terminal experience but lack push notifications, clean UIs for answering questions or viewing git diffs, and easy setup.

We wanted our integration to fully mirror the native Claude Code experience, including terminal output, permissions, notifications, and mode switching. The Claude Code SDK and hooks don't support all of this, so we made a CLI wrapper that parses the session file at ~/.claude/projects and the terminal output to capture user and agent messages. We send these messages to our platform, where they're displayed in the web and mobile apps in real time via SSE. Our CLI wrapper monitors for input from both the Omnara platform and the Claude Code CLI, continuing execution when the user responds from either location. Our entire backend is open source: https://github.com/omnara-ai/omnara.

Omnara isn't just for Claude Code. It's a general framework for any AI agent to send messages and push notifications to humans when they need input. For example, I've been using it as a human-in-the-loop node in n8n workflows for replying to emails. But every Claude Code user we show it to gets excited about that application specifically so that’s why we’re launching that first :)

Omnara is free for up to 10 agent sessions per month, then $9/month for unlimited sessions. Looking forward to your feedback and hearing your thoughts and comments!

Hi HN, I’ve been working on Gore – a port of the classic Doom engine written in pure Go, based on a ccgo C-to-Go translation of Doom Generic. It loads original WAD files, uses a software renderer (no SDL or CGO, or Go dependencies outside the standard library). Still has a bit of unsafe code that I'm trying to get rid of, and various other caveats.

In the examples is a terminal-based renderer, which is entertaining, even though it's very hard to play with terminal-style input/output.

The goal is a clean, cross-platform, Go-native take on the Doom engine – fun to hack on, easy to read, and portable.

Code and instructions are at https://github.com/AndreRenaud/Gore

Would love feedback or thoughts.

Today I’m releasing the XR2000: A programming challenge with extensive science fiction backstory.

Hi everyone,

I've made an open-source library for using design by contract (DbC) principles in the Zig programming language.

It's called Zig-DbC, and it currently provides the following features:

- A simple API to define preconditions, postconditions, and invariants.

- Contracts are active in `Debug`, `ReleaseSafe`, and `ReleaseSmall` modes to catch bugs early.

- All checks are removed at compile time in `ReleaseFast` mode for zero performance cost.

- An optional mode to handle partial state changes in functions that return errors.

- Transparent error handling that propagates errors from your code to the caller.

Project's GitHub repo: https://github.com/habedi/zig-dbc

After absorbing opinions from the previous post, I wrote an article specifically introducing what a Spatial Web Browser Engine is.

I was working on validating some of my own project ideas. While trying to find how to validate my idea, I realized the process itself could be turned into a tool.

A few late nights later, I had something that takes any startup idea, fetches discussions, summarizes sentiment, and gives a quick “validation score.”

It’s very rough, but it works, and it’s already making me rethink a few of my own ideas.

It's still a work in progress. I don't actually know what I'm doing, but I know it's worth it. Honest feedback welcomed! Live demo here: https://validationly.com/

I built an open-source AI agent for security testing to find and fix vulnerabilities in your code.

I’ve noticed how bad security vulnerabilities have gotten with everyone shipping AI code slop, so I wanted to build something that allows for vibe-coding at full speed without compromising security.

Traditional security tools aren’t effective, and manual pen-testing can’t keep up with the rapidly growing AI code

This tool runs your code dynamically, finds vulnerabilities, and validates them through actual exploitation.

You can either run it against your codebase or enter your (or someone else’s) domain to scan for vulnerabilities.

Good luck, have fun, hack responsibly!

Hello HN! Recently, we have released Nocturne 3.0.0, which is a complete replacement for the (now unusable) Spotify Car Thing stock firmware. We're proud to eliminate more e-waste in the world.

# Changes from v2 - Bluetooth tethering for car use (no more Raspberry Pi in the car) - Full graphics acceleration - Native Spotify login (no more client ID/secret) - Start DJ from the Car Thing - Podcast support - Gesture control - New settings - Boot to Now Playing - Spotify Connect device switcher - Support for Japanese, Simplified Chinese, Traditional Chinese, Korean, Arabic, Devanagari, Hebrew, Bengali, Tamil, Thai, Cyrillic, Vietnamese, and Greek - Full knob control support - Local file support - Preset button support - Status bar on home (shows time & Bluetooth/Wi-Fi) - Auto brightness - Hold settings button for power menu - Lock screen showing time full screen (press settings button) - DJ preset binding (hold preset button while DJ is playing in Now Playing) - Spotify mixes in Radio tab (Discover Weekly, daily mixes, etc.) - OTA updates - + MUCH more (this is just the important stuff!)

# Flashing A guide to flashing Nocturne 3.0.0 is in the README. Bluetooth will work out of the box, or choose an alternative in the Setting up Network section. Hotspot capability from your phone and plan are required for Bluetooth.

# Notes This wouldn’t be possible without our donors and the rest of the Nocturne Team. We hope you’ll enjoy it, as we've spent thousands of hours working on it!

Consider buying the team a coffee if you can https://usenocturne.com/support

https://github.com/usenocturne/nocturne/releases/tag/v3.0.0

A motion-graphic comparison website in the vein of LMArena. The videos are rendered via Remotion.

We hope that AI will be used in interesting ways to help with video production, so we wanted to give some of the models available today a shot at some basic graphics.

Built an iOS app that runs a local OCR server using Apple's Vision Framework.

Creates a REST API endpoint accessible from any device on your network. No cloud services needed - everything processes locally on the phone.

Available on App Store (searching "OCR Server").

Would appreciate feedback on the architecture or similar mobile-as-server projects you've seen.

Hey there, I'm a solo dev behind KewlTools.

So I'm one of those people who like to build their own tools/utilities whenever they a) want to get something done, or b) want to learn something. c) don't want to spend ages finding a utilty and going through walls of ads.

I gradually built 66 tools, all free, fast, ad-free, and zero login.

I (and my family/friends) use these daily, so thought some of you will find it useful!

Any suggestions, improving current tools or adding new ones, please let me know.

I'm building a manually curated catalog of games made by the HN community.

I'v found several small gems while browsing HN, and it would've been quite difficult to find those games otherwise. So I've started a personal quest to collect those games into a public catalog with open source code and open data (everything is available on GitHub: https://github.com/labarilem/hn-games).

Currently the collection has data up to the end of 2022. The plan, of course, is to gather all data up to today and then keep updating the collection.

I've published this early version of the catalog to gather some insights and feedback if possible. Especially about UI/UX, which has been mostly LLM-coded to speed things up (I'm not an expert in those domains).

You can browse the catalog at these addresses:

- https://hackernews.games/

- https://hn-games.marcolabarile.me/

Please let me know what you think about the project!

I'm building a workspace for doing focus work while also managing constant multi-tasking and context-switching that's required in a lot of roles. Currently this is done by using a notes system where we can split-screen AI chat, an object-based data table system for tracking entities, as well as viewing other documents.

We can define agents directly from documents by giving instructions in plain language and then defining the trigger conditions. This helps automate workflows directly in the place where we're reading and writing content.

I'm continuing to experiment with instructions and templates to figure out the best ways to automate tasks like content creation and responding to emails or leads.

Hey HN, I’m Julia, my team and I are building Rag-in-a-Box (https://www.joinable.ai/rag-in-a-box), hosted RAG service that let’s builders of any skill level launch their own RAG app loaded with their own data in minutes.

[ What can you do ] 1. Load your documents (PDFs, CSV, PPTs, Word Docs, etc) and make them searchable instantly. All your data stays private and encrypted.

2. Choose latest open source LLM (Llama 4, Deepseek, GPT-oss, etc) to interact with your docs

3. Access your hosted RAG via API - build your own custom front end or integrate with your existing product / service.

[ Why we built it ] 1. I love Notebook LM, but it has limitations: - you stuck with using Gemini models - no API - I wanted to integrate my notebook with my other apps but couldn't

2. Prior to this my teammate Brian was Head of AI at a public software company, teams across the organization constantly asked him to build various RAG apps for internal use (searching marketing docs, pulling financial information, etc)

But these projects weren’t a company priority, required lots of resources, therefore moved slowly through approvals... and were scrapped at the end.

This is why we decided to build RAG-in-a-Box to make it easy for anyone build their own RAG app.

[ How does it work ] 1. Create a Data Collection (“folder”) this will become your mini RAG application

2. Dump your data

3. Click Launch - your RAG is ready to Go

4. Use Joinable dashboard to start interacting with your RAG app or

5. Connect to your RAG App via API - add data, delete docs, generate markdown responses, pick LLM models, etc all available through API

Optional - embed your RAG App into your existing application or build custom front end or even a mobile app

Here are some quick tutorials to get started: https://joinable.gitbook.io/joinable-api-docs

We’ve just launched and would love you to try it and hear your feedback!

Hi everyone. I have been learning about deep learning for some time, and I've tried to implement CNN, neural networks, U-Net, transformers etc. to learn and understand them more and also to get my hands dirty on the frameworks, however I've noticed that many tutorials online are not very detailed, so concepts are not explained clearly, so people would understand neural networks only shallowly. On the other hand, many sources like books may show many, many equations but do not show the main points, so people reading those books would get lost in mathematical details, which hampers learning. When I tried reading about RNN, or LSTM, I've noticed that many tutorials do not fully explain them. Some show pictures to make visualization easier, some show forward equations but the backward equations are not discussed. But there is something which I don't think is talked about much, is that many tutorials, even if they show the backpropagation, only limit it to a single RNN layer (this is also true for LSTM/GRU).

Some time ago, I read this book called "Neural Network design" by M. Hagan, and I found the explanations of the book to be quite good (even though the book is not new). The book explains things clearly enough for you to build everything and does not handwaive anything. When I checked the part about RNN, I noticed that the book explains how to do backprop for RNN with arbitrary connections, not just one RNN layer, which I think is something not many sources online show. The book also derives for the conditions of different delays, which I think is completely skipped in other sources.

So I decided to go ahead and implement it. The URL provides link to my implementation, which includes: The implementation includes:

- Full BPTT for RNN networks with arbitrary recurrent connections and delays

- Comprehensive gradient checking using finite differences

- Bayesian regularization and multiple optimization algorithms

- Extensive numerical validation throughout

I think I learned a lot during the implementation, both about how to implement a neural network, and also about how to structure my program, etc. I tried to be systematic and included tests for correctness of backprop by approximate difference equation (you know the [f(x+delta)-f(x-delta)]/(2*delta) thing). This also made me try to learn about Einstein summation (using Numpy) which really help things. During this period, I also learned that equation (14.39) has some slight error which is fixed in later equations (this was confirmed in private emails with the authors). The gradient checking was essential for debugging these subtle mathematical issues.

Key lessons:

- Systematic software development techniques, coupled with mathematical rigour, help catch ML bugs more effectively.

- Implementing from first principles help solidify your understanding and reveal the inner workings which frameworks hide.

- Einstein summation is a good thing to make the maths much cleaner.

Even though RNN is not the latest architecture, I think there is value in firstly grounding in fundamentals before jumping to more complex models.

Our little team built a small, open‑source browser extension because I wanted page summary/translate/chat without sending data off‑device.

It runs gpt‑oss locally via Ollama. Optional: switch to GPT‑5 if you add an OpenAI API key.

Try: 1.Install the extension (GitHub). 2.Install Ollama, then ollama pull gpt-oss. 3.In NativeMind choose “gpt‑oss”, then highlight text or click the toolbar button to summarize/translate/chat.

Notes: no backend servers from us; with gpt‑oss everything stays on your machine. With GPT‑5, requests go to OpenAI.

Would love feedback on setup friction, performance on your hardware, and missing features. Issues/PRs welcome.

I’ve released sarpro, an open-source Rust tool for converting Sentinel-1 Synthetic Aperture Radar (SAR) Ground Range Detected (GRD) products into GeoTIFF or JPEG.

Why? You know batch conversion can be painfully slow. sarpro is fast (Rust-based), supports CLI, GUI, and Rust API, and includes: • Batch processing for many GRDs at once • Synthetic RGB composition from polarization channels • Autoscaling and padding for consistent image rendering • Output as GeoTIFF or JPEG

On my machine M4Pro12, (~500MP) dual-band image scaled to 2048px on the long side and carrying metadata took just 35 seconds with CPU < 22% usage.

I’d love feedback from the remote sensing / EO community — ideas for more processing modes, more RGB presets, or cloud pipeline integration with other tools and workflows welcome.

My brother and I decided to make a word game. We wanted something fun that could be generated on a daily basis but with the amount of challenge it takes to finish a cup of coffee.

Looking for any feedback or thoughts, and please enjoy!

Hey everyone! I built this because I was tired of manually handling customer support for my web app and couldn't get an AI system to handle requests reliably.

I tried different AI tools to help with support tickets, but when they handled requests incorrectly, it was impossible to determine why, and even harder to figure out what I needed to change to improve the system.

I wanted to break down my logic of how the AI should think through the problem step-by-step, but everything had to be crammed into one prompt. Without the volume of clean training data needed for fine-tuning, I was stuck with prompt engineering guesswork.

What Chainix does: You drag and drop steps into a visual flowchart. Each step gets its own inference instructions, and based on the output, it branches to different next steps. The AI can also pause mid-flow to call your functions or check variables, then continue. This lets you visually map out exactly how you want the AI to think through the problem (like a flowchart).

I built it with flexibility in mind - you can create something as simple as a two-step workflow or build complex custom logic with multiple branches and conditions.

The key: when something goes wrong, you can see exactly which step failed. Instead of one big black box, you have a chain of smaller, debuggable pieces. My support flow might classify the ticket, look up account info, check for known issues, then craft a response. When the AI did something wrong, I could see "oh, this step classified the ticket incorrectly" and just fix that inference step (or add a new one).

It's handling ~60% of my support requests reliably now (and correctly ignoring the rest), so I'm very happy with it! The biggest win is that I can actually see how the AI is reasoning through each step, so fixing issues is straightforward instead of guesswork.

This works for any workflow involving text interpretation and action - content moderation, document processing, lead qualification, etc.

You can try it at https://www.chainix.ai - would love to hear if other people have hit this same wall with AI tools! Also curious what other workflows people might want to build with this approach.

Bulk export Twitter/X posts to CSV, JSON, or Excel with Twitter Post Exporter. Quick, easy social media backup & analytics.

Hey all, I made a daily logic puzzle as a passion project, and I'd love for you to test it.

If you are interested, read on for a bit of a devlog and backstory.

I have always been obsessed with logical deduction puzzles. These are my actual logic puzzle books from when I was 10 years old: https://i.imgur.com/h1HULLk.png I would spend hours doing these puzzles, and it's been on my backlog for a long time to make something that captures that feeling again.

This particular kind of grid-based logic puzzle is sometimes referred to as "Einstein's Riddle" or "a zebra puzzle". There are other online implementations of these puzzles but I personally either disliked the UX or disliked the format of the clues. For Griddle I opted for a more abstracted format, using emoji-style SVGs. So, if the categories are 'job' and 'pet', a simple clue like [Astronaut is Dog] would mean that The job 'Astronaut' has a one-to-one relationship with 'Dog' from the pet category.

Getting it to play nicely on mobile was a tricky beast, because the standard format of this kind of puzzle is a choppy grid that is as wide as it is tall, wherein every category is shown to have a relation to every other category. To solve this, I added two UX options: 'full view' and 'simple view'. In simple view, your categorical relations are presented as discrete square grids in a single column. This is very nice on mobile, but lacks the visual link that cues you into make transitive conclusions. If you opt for full view, you have to scroll left and right, but you get the full visual context of how things relate to each other transitively. The column and row headers are 'sticky' and will follow you so you don't forget the context of the cell you're looking at.

My favorite part of developing this was figuring out the logic for generating interesting, solvable puzzles with a minimum number of clues. Generating the puzzle solution matrix is easy (and all randomization is done via deterministic/seeded shuffle with the puzzle date as the seed, so re-generations are consistent). Select a number of categories, select a number of items in each, and map them uniquely one-one-one.

I defined a a number of clue types that have their own logic.

equality: A is B inequality: A is not B 2-category XOR: A is either B or C (B and C are from same category) 3-category XOR: A is either B or C (B and C are from different categories) 2-category double negative: A is not B or C (B and C are from same category) 3-category double negative: A is not B or C (B and C are from different categories) 2-category group: A and B are paired with C and D (A/B share a category; C/D share another) 3-category group: A and B (same category) are paired with C and D (from two different categories), in some order 4-category group: A (cat1) and B (cat2) are paired with C (cat3) and D (cat4), in some order 3-category biconditional (same item, both true): A is B iff A is C 3-category biconditional (same item, both false): A is B iff A is C 3-category biconditional (different items, both true): A is C iff B is C 3-category biconditional (different items, both false): A is C iff B is C 4-category biconditional (both true): A is C iff B is D 4-category biconditional (both false): A is C iff B is D

Then the solver algorithm does all the work.

- generate a candidate clue of each allowed type - filter out invalid clues - weight the 'impact' of each clue based on how many cells it would fill - select one with weighted (and seeded) randomization - but don't exclude clues with weight 0, because sometimes a clue is only useful later! - iterate through all transitive and deductive logic and all past clues until no more cells are filled - if not solved, do it again until solved - randomize clue order - run the solver AGAIN, this time trimming out any that don't contribute to puzzle progress

There is probably a more efficient and smarter way to do this, but it works.

Create professional invoices in seconds, no signup required. Clean interface, multiple templates, PDF export, and API access. Built for developers and freelancers who value simplicity and privacy.