🛠️ Hacker News Tools

[Other] Show HN: Nibble An attempt at a single pass LLVM frontend in ~3000 lines of C without external dependencies, malloc, or an AST. Included are some graphical examples. The IR isn't perfect, and the README touches on one particular downfall

[API/SDK] Show HN: Claude-pee: use Claude -p without the programmatic usage credit pool Anthropic announced today that starting June 15, paid Claude plans get a separate monthly credit pool for programmatic usage (claude -p, Agent SDK). Seems OK at first glance but it turns out the monthly credit pool is charged at API rates, which effectively kills any serious programmatic usage for hobbyists.<p>This is a small Claude Code wrapper which runs claude in a PTY, injecting input, finding the session transcript jsonl file, and using a stop hook to determine when claude is done.<p>It&#x27;s a drop-in replacement for claude code. All arguments except the &quot;-p&quot; argument are forwarded as-is.<p>claude-pee -p &quot;hello world&quot;<p>Could be used as a cheap way of using Anthropic subscriptions for OpenClaw, Hermes, etc.<p>Written in Rust, MIT license.

[Other] A SQL-Inspired Query Language Designed for Event Sourcing (2025)

[Other] Show HN: Neural window manager, neural network moving windows from mouse actions I&#x27;d been mulling over this crazy idea for a while. Can programs be generated? Inspired by recent advances in world models, I wondered if we could do away with source code and generate pixels directly and interactively.<p>As an experiment to answer this, I set out to create a neural window manager, training a neural network to predict what the screen would look like next.<p>Basically, the idea was to generate the next frame based on the last two frames and the mouse position. That&#x27;s it: moving windows without programming an event system, just a simple convolutional neural network guessing pixels.<p>To implement the experiment, I used Pygame to simulate a turquoise desktop background, a gray window with a navy blue title bar, a white cursor, and four colors in total. Then, a bot randomly dragged the window, and I recorded everything, processing the frames as color index matrices (not RGB, to avoid complications) and the mouse delta (dx, dy, click) that caused each transition. 8000 frames, a few minutes in Colab.<p>The model is a unitary neural network (UNET). The encoder compresses the stacked frames, the decoder reconstructs the next one, and the mouse vector coordinates are projected with a linear layer to fit the spatial size of the bottleneck. There, they are concatenated before decoding, so that motion information feeds each jump connection.<p>And it works! Which still surprises me a little. You can drag, and the window follows you; when you release, it stops. There&#x27;s no internal state, no (x, y) coordinates anywhere. The model infers the position from what it sees, which works until it doesn&#x27;t. But after a couple of seconds of strange movement, the window starts to distort.<p>This will probably improve with more computing power for training and more examples, but to narrow the scope of the experiment and test it within a web browser, I decided to abandon the rendering aspect and have the model predict primitives instead of pixels, simply converting the motion engine into a neural network.<p>Basically, I trained a small MLP to receive (distance to the title bar, distance to the resize point, click) and generate (dx, dy, dw, dh), with two separate heads: one for moving and one for resizing. The trick is that they share nothing except the click signal, so the model can&#x27;t confuse dragging with resizing. I then exported it to ONNX as well, and now everything runs in the browser, without a server, just a canvas element and two small neural networks communicating with each other.<p>With this new approach, the renderer remains deterministic, with rectangles drawn in JavaScript, but the window&#x27;s behavior (where it moves, how it resizes) is learned from examples. It feels like a peculiar middle ground between traditional and neural, so you can feel the space the network has learned by interacting with it: dragging near the title bar moves it, but approaching the corner resizes the window. There are no conditionals or hitbox code; the network simply learned where those areas are from examples.<p>Sometimes it gets confused near the edges, which, frankly, is more interesting than if it worked perfectly; you can perceive how the probability changes. This makes sense when you think about it, because no (x, y) coordinates are stored in these models; the position is implied in the activations. It works well for short sequences, but fails when asked to maintain state over time.<p>Update: A few weeks later, Meta published the Neural Computers article (2604.06425, it&#x27;s worth reading). The premise is the same, but they go much further: cli and uis, real programs. Their failure modes are practically identical to those I found with the pure pixel version: &quot;challenges persist with routine reuse, controlled updates, and symbolic stability.&quot; which is a fancy way of saying that the window blurs after a few seconds (that was the reason for choosing deterministic rendering).

[Database] Launch HN: Ardent (YC P26) – Postgres sandboxes in seconds with zero migration Hey HN! We’re Vikram and Evan from Ardent (<a href="https:&#x2F;&#x2F;tryardent.com">https:&#x2F;&#x2F;tryardent.com</a>). We&#x27;re building database sandboxes for you and your coding agents.<p>In the last two years coding agents have gotten dramatically more capable at handling complex engineering tasks. But without access to a realistic sandbox at the DB layer for testing, they ship garbage that can take down production databases. I spent over a year building an AI Data Engineer that failed for this exact reason. Evan spent the last 12 years in data engineering and hit this wall building agents at his last company.<p>Ardent was built to make it possible for coding agents to get near instant access to production-like sandboxes so they can test their work. To do this we write a replication stream out of the target DB, scaling with kafka onto a read replica with copy on write enabled and autoscaling compute (we currently prefer neon as a primary branching engine due to their implementation of these properties).<p>Our replication stream uses logical replication + ddl triggers to enable usage on any hosted postgres DB since most platforms do not allow physical replication which is traditionally used for creating replicas.<p>This provides a few primary benefits:<p>1. Does not require a platform migration to a DB provider like neon, allowing strong separation of production and development concerns. 2. Minimal impact on the production database while allowing clones to spin up in &lt;6s, even at TB scale with copy-on-write<p>Security matters a lot with cloning production so we run a proxy layer to generate custom postgres URLs and route all connections to allow more granular access control to clones, prevent credential leak, and follow a split plane architecture to allow full data residency on your cloud through BYOC.<p>We also support anonymization through the ability to register SQL that runs on branches before they are returned. This has been used for PII redaction and branch modification.<p>Our goal is to make every data infrastructure platform “cloneable” in one place so agents can fully test the impact of their changes on production like data environments without risk.<p>Here&#x27;s a demo of it: <a href="https:&#x2F;&#x2F;youtu.be&#x2F;5S1kwPtiRU0" rel="nofollow">https:&#x2F;&#x2F;youtu.be&#x2F;5S1kwPtiRU0</a><p>We’d love to understand how you work with coding agents on the DB and if you try Ardent (it&#x27;s free to get started) what worked, what broke and what’s missing.

[DevOps] Show HN: Mistle – Open-source infrastructure for running sandboxed coding agents Hi HN, I&#x27;m Jonathan. My co-founder, Thomas, and I started building Mistle in Feb.<p>We saw larger tech companies like Ramp (Inspect) and Stripe (Minions) build this internally and thought an open source version should exist.<p>We made a few very intentional decisions when working on this:<p>1. Credentials are kept out of the sandbox. Authorized access goes through a proxy, so agents do not directly receive credentials.<p>2. The harness is not our problem. We&#x27;re not going to tackle things like memory, self-learning.<p>3. No magic. Configurations are explicit. You can bring your own keys for models, sandboxes, and other providers. You can write your own instructions and agent.<p>Mistle can be run locally with a single command: <a href="https:&#x2F;&#x2F;github.com&#x2F;mistlehq&#x2F;mistle#run-mistle-locally" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;mistlehq&#x2F;mistle#run-mistle-locally</a><p>Questions, feedback and ideas are welcome!

[Other] Show HN: Promptcellar – capture every Claude Code prompt as JSONL in your repo

[Other] Show HN: Rotunda - A browser built for agents with simulated typing Hi HN! Pierce here.<p>Rotunda is a firefox fork primarily intended for agent use, which I’ve been hacking on nights&#x2F;weekends.<p>There was a [lengthy](<a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=48024859">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=48024859</a>) discussion last week on how expensive computer use models are. The cost is going to drop eventually, but I think on some level it&#x27;s still usually the wrong primitive. The web gives us access to beautiful structured formats, plaintext, etc... why throw that away if we don&#x27;t have to?<p>I realized at some point that for 99% of automations I just want agents to be able to control my Chrome instance. But that’s easier said that done: CDP (the Chrome automation protocol) leaks a ton of state about being programmatically controlled, either by toggling window attributes or by running `page.evaluate()` commands right in the page context. Plus if you look at an automation running it&#x27;s pretty obvious what happens: the mouse jumps around, fields are filled instantly, etc.<p>Rotunda tries to fix this. Its standout features:<p>- Realistic simulation of mouse movements and keyboard commands, powered by a trained RNN on my own timing patterns from the last week. (still feel weird about opting-in to a key logger but whatever)<p>- Doesn’t lie about its host specs, only fibs about some client side details. Stealth browsers are too easy to flag statistically when you’re adding noise to canvas pixels or audio pipelines.<p>- It runs on your local device with a CLI or Playwright API accessible to Claude, Codex, or whatever your harness-de-jure today looks like.<p>- Patches modern Firefox (150) with an agentic harness to keep this updated over time<p>MPL-2.0 on GitHub: <a href="https:&#x2F;&#x2F;github.com&#x2F;monkeysee-ai&#x2F;rotunda" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;monkeysee-ai&#x2F;rotunda</a><p>Longer writeup on the design choices: <a href="https:&#x2F;&#x2F;pierce.dev&#x2F;notes&#x2F;a-browser-for-agents" rel="nofollow">https:&#x2F;&#x2F;pierce.dev&#x2F;notes&#x2F;a-browser-for-agents</a><p>Also check out the demo on the site! <a href="https:&#x2F;&#x2F;www.rotunda.sh&#x2F;" rel="nofollow">https:&#x2F;&#x2F;www.rotunda.sh&#x2F;</a><p>Pretty excited by how this turned out but we’re still super early. Give it a try and please flag any issues!

[Other] Why I'm leaving GitHub for Forgejo

[CLI Tool] Show HN: FixMyNPM, CLI to fix your insecure npm config With recent issues with npm configuration issues it&#x27;s essential that we have a tool like this to fix basic issues

[Code Quality] Show HN: Gox – Strict static analyzer for Go designed for LLM-written code

[DevOps] Show HN: Kunobi 1.0 – A local-first Kubernetes workspace for GitOps teams

[Monitoring/Observability] Show HN: Torrix, self hosted, LLM Observability,(no Postgres, no Redis) I work as a SAP Integration consultant and built this as a side project. Friction point: Most self hosted LLM observability tools require Postgres, Redis and non trivial infrastructure. Teams just want to see what their agents are actually doing in Production, that set up cost discorages adoption. Torrix runs as a single docker contained backed by SQLite. The full install is:<p>curl -o docker-compose.yml <a href="https:&#x2F;&#x2F;raw.githubusercontent.com&#x2F;torrix-ai&#x2F;install&#x2F;main&#x2F;doc" rel="nofollow">https:&#x2F;&#x2F;raw.githubusercontent.com&#x2F;torrix-ai&#x2F;install&#x2F;main&#x2F;doc</a>... docker compose up<p>No external dependencies. All data stays in a local SQLite file on your machine.<p>It logs LLM calls through a HTTP proxy or a python&#x2F;Node SDK : tokens, cost, latency, full prompt and response traces, reasoning token capture. Works with OpenAI, Anthropic, Gemini, Groq, Mistral, Azure Open AI and any Apen AI compatible end point.<p>Things I added as I actually used it on real agent pipelines: cost forecasting and hard budget caps, PII masking, model routing rules, evals with golden runs, AI judge, a prompt library with version history, run tags for filtering by environment, MCP server so AI Assistants can query your own logs and OTLP&#x2F;HTTP ingestion for apps aöready using OpenTelemetry.<p>Community edition is free for one user with 7-day retention. Pro adds teams, RBAC, 30 day retention, API key management, full text search and audit logs.<p>SQLite doesn&#x27;t scale to high write throughput. This is aimed at teams logging hundreds to low thousands of LLM calls per day, not millions. Happy to hear what people think and what is missing.<p>GitHub &#x2F; install: <a href="https:&#x2F;&#x2F;github.com&#x2F;torrix-ai&#x2F;install" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;torrix-ai&#x2F;install</a> Website: <a href="https:&#x2F;&#x2F;www.torrix.ai" rel="nofollow">https:&#x2F;&#x2F;www.torrix.ai</a>

[Other] GitHub Actions issued GitHub_TOKEN disclosure in GitHub Actions logs

[Database] Show HN: Monghoul – Desktop MongoDB GUI with schema-aware autocomplete and MCP Last year I decided to start a fun side project - a love child of VS Code and NoSQLBooster.<p>I wanted a GUI that looks modern and snappy, minimal, not like 2003 MS Excel with dozens of buttons and dropdowns everywhere. I also wanted it to have a smart autocomplete that actually knows a schema, not just keys of the current collection, but their types and enum values. I wanted to type find({status: &quot;}) and see &quot;pending&quot;, &quot;active&quot;, &quot;cancelled&quot; in the autocomplete suggestions.<p>As a tech stack, I chose Tauri for the shell, Bun for the sidecar running the MongoDB driver and a tRPC server, and react, tailwind, react-query for the UI. The installer is around 33 MB.<p>Also it has a built in MCP server that allows your AI tools to fully control the app: write queries, build charts, organize workspace, find and restore tabs that you once closed etc.<p>Using the combination of tauri + bun sidecar + trpc with react-query was the best decision: - startup under 2 seconds - end-to-end type safety without a need to update client interfaces on back-end changes - client optimistic updates are super easy to do, so everything feels instant

[Other] Recreation of the 1956 IPL-I version of the Logic Theorist theorem prover

[Other] Zero-native – Build native desktop apps with web UI

[Other] Show HN: Needle: We Distilled Gemini Tool Calling into a 26M Model Hey HN, Henry here from Cactus. We open-sourced Needle, a 26M parameter function-calling (tool use) model. It runs at 6000 tok&#x2F;s prefill and 1200 tok&#x2F;s decode on consumer devices.<p>We were always frustrated by the little effort made towards building agentic models that run on budget phones, so we conducted investigations that led to an observation: agentic experiences are built upon tool calling, and massive models are overkill for it. Tool calling is fundamentally retrieval-and-assembly (match query to tool name, extract argument values, emit JSON), not reasoning. Cross-attention is the right primitive for this, and FFN parameters are wasted at this scale.<p>Simple Attention Networks: the entire model is just attention and gating, no MLPs anywhere. Needle is an experimental run for single-shot function calling for consumer devices (phones, watches, glasses...).<p>Training: - Pretrained on 200B tokens across 16 TPU v6e (27 hours) - Post-trained on 2B tokens of synthesized function-calling data (45 minutes) - Dataset synthesized via Gemini with 15 tool categories (timers, messaging, navigation, smart home, etc.)<p>You can test it right now and finetune on your Mac&#x2F;PC: <a href="https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;needle" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;needle</a><p>The full writeup on the architecture is here: <a href="https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;needle&#x2F;blob&#x2F;main&#x2F;docs&#x2F;simple_attention_networks.md" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;needle&#x2F;blob&#x2F;main&#x2F;docs&#x2F;simp...</a><p>We found that the &quot;no FFN&quot; finding generalizes beyond function calling to any task where the model has access to external structured knowledge (RAG, tool use, retrieval-augmented generation). The model doesn&#x27;t need to memorize facts in FFN weights if the facts are provided in the input. Experimental results to published.<p>While it beats FunctionGemma-270M, Qwen-0.6B, Granite-350M, LFM2.5-350M on single-shot function calling, those models have more scope&#x2F;capacity and excel in conversational settings. We encourage you to test on your own tools via the playground and finetune accordingly.<p>This is part of our broader work on Cactus (<a href="https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;cactus" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;cactus</a>), an inference engine built from scratch for mobile, wearables and custom hardware. We wrote about Cactus here previously: <a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=44524544">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=44524544</a><p>Everything is MIT licensed. Weights: <a href="https:&#x2F;&#x2F;huggingface.co&#x2F;Cactus-Compute&#x2F;needle" rel="nofollow">https:&#x2F;&#x2F;huggingface.co&#x2F;Cactus-Compute&#x2F;needle</a> GitHub: <a href="https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;needle" rel="nofollow">https:&#x2F;&#x2F;github.com&#x2F;cactus-compute&#x2F;needle</a>

[Database] Quack: The DuckDB Client-Server Protocol

[IDE/Editor] Show HN: Agentic interface for mainframes and COBOL Hi HN, we’re Sai and Aayush, and we’re building Hypercubic (<a href="https:&#x2F;&#x2F;www.hypercubic.ai&#x2F;">https:&#x2F;&#x2F;www.hypercubic.ai&#x2F;</a>), bringing AI tools to the mainframe and COBOL world. (We did a Launch HN last year: <a href="https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=45877517">https:&#x2F;&#x2F;news.ycombinator.com&#x2F;item?id=45877517</a>.) Today we’re launching Hopper, an agentic development environment for mainframes.<p>You can download it here: <a href="https:&#x2F;&#x2F;www.hypercubic.ai&#x2F;hopper">https:&#x2F;&#x2F;www.hypercubic.ai&#x2F;hopper</a>, and you can also request access and immediately get a mainframe user account to play with.<p>There&#x27;s also a video runthrough at <a href="https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=q81L5DcfBvE" rel="nofollow">https:&#x2F;&#x2F;www.youtube.com&#x2F;watch?v=q81L5DcfBvE</a>.<p>Mainframes still run a surprising amount of critical infrastructure: banking, payments, insurance, airlines, government programs, logistics, and core operations at large institutions. Many of these systems are decades old, but they continue to process enormous transaction volumes because they are reliable, secure, and deeply embedded into business operations.<p>A lot of that software is written in COBOL and runs on IBM z&#x2F;OS. The development environment looks very different from modern cloud or Unix-style development. Instead of GitHub, shell commands, package managers, and CI pipelines, developers often work through TN3270 terminal sessions, ISPF panels, partitioned datasets, JCL, JES queues, spool output, return codes, VSAM files, CICS transactions, and shop-specific conventions.<p>TN3270 is the terminal interface used to interact with many IBM mainframe systems. ISPF is the menu and panel system developers use inside that terminal to browse datasets, edit source, submit jobs, and inspect output. It is powerful and reliable, but it was designed for expert humans navigating screens, function keys, and fixed-width workflows, not AI agents.<p>A simple COBOL change might require finding the right source member, checking copybooks, locating compile JCL, submitting a job, reading JES&#x2F;SYSPRINT output, interpreting condition codes, patching fixed-width source, and resubmitting.<p>Much of this work is so well-defined and repetitive that it&#x27;s a good fit for agentic AI. To get that working, however, a chatbot next to a terminal is not enough. The agent needs to operate inside the mainframe environment.<p>Hopper combines three things: (1) A real TN3270 terminal, (2) Mainframe-aware panels for datasets, members, jobs, and spool output, and (3) An AI agent that can operate across those z&#x2F;OS surfaces.<p>For example, here is a tiny version of the kind of thing Hopper can help debug:<p><pre><code> COBOL: IDENTIFICATION DIVISION. PROGRAM-ID. PAYCALC. DATA DIVISION. WORKING-STORAGE SECTION. 01 CUSTOMER-BALANCE PIC 9(7)V99. PROCEDURE DIVISION. ADD 100.00 TO CUSTOMER-BALNCE DISPLAY &quot;UPDATED BALANCE: &quot; CUSTOMER-BALANCE STOP RUN. JCL: &#x2F;&#x2F;PAYCOMP JOB (ACCT),&#x27;COMPILE&#x27;,CLASS=A,MSGCLASS=X &#x2F;&#x2F;COBOL EXEC IGYWCL [&#x2F;&#x2F;COBOL.SYSIN](https:&#x2F;&#x2F;cobol.sysin&#x2F;) DD DSN=USER1.APP.COBOL(PAYCALC),DISP=SHR [&#x2F;&#x2F;LKED.SYSLMOD](https:&#x2F;&#x2F;lked.syslmod&#x2F;) DD DSN=USER1.APP.LOAD(PAYCALC),DISP=SHR </code></pre> A human would submit this job, inspect JES output, open `SYSPRINT`, find the undefined `CUSTOMER-BALNCE`, map it back to the source, patch the member, and resubmit. Hopper is designed to let an agent operate through that same loop autonomously.<p>Hopper is not trying to hide the mainframe behind a generic abstraction, and it&#x27;s not a chatbot. The design principle is simple: preserve the fidelity of the mainframe environment, but make it accessible to AI agents.<p>Sensitive operations require approval, and the terminal remains visible at all times.<p>Once agents can operate inside the mainframe environment, new workflows become possible: faster job debugging, automated documentation, safer code changes, test generation, migration planning, traffic replay, and modernization verification.<p>We’re curious to hear your thoughts! especially from anyone who has worked with mainframes, COBOL or has done legacy enterprise modernization.