A new wave of AI tooling appeared on Hacker News over the last six weeks. Most of what shows up there is a demo, a benchmark or a joke project, so it is worth separating the launches that are real products from the ones that are weekend experiments. The eight tools below are the ones that are actually usable today: either an open-source repository you can run or a hosted product you can sign up for. The mix leans toward developer tooling, with a handful of consumer apps that solve one specific annoyance rather than promising to change everything.
Janus — one binary for every GPU
Local inference has a familiar frustration: the runner you need depends on the graphics card you happen to own. Janus is a Go binary that runs GGUF models through the Vulkan API, so the same executable works across AMD, Intel and Nvidia hardware. Instead of maintaining a separate build for each vendor, you ship one binary and let Vulkan handle the translation. The model weights stay on your machine, so prompts and data do not go to a hosted API. It is aimed at people who want local inference without committing to one GPU ecosystem, and the setup still expects you to be comfortable with drivers and command-line tooling.
Engrim — memory that outlives the session
If you run more than one AI coding assistant, you have probably noticed that none of them remember anything between sessions. Engrim is a local-first memory engine built on SQLite that gives AI command-line tools a shared, persistent place to store context. It is described as universal, meaning one memory store can back different coding agents and CLIs rather than locking you into a single vendor. Because everything lives in a SQLite file on your own disk, the history stays local and portable. It is a direct answer to a problem that gets worse the more agents you use, and it leaves you responsible for managing that file as it grows.
Offrun — one workspace for all your agents
Running two or three coding agents at once usually means three terminal windows and no clear picture of what any of them is doing. Offrun is a workspace that sits above the agents you already use and lets you manage them from a single interface. It is not an agent itself; it coordinates the ones you already configure and pay for. The value shows up once you have more than one agent running, since the alternative is constant context switching between windows and dashboards. Pricing and limits were not stated clearly at launch, and you still configure each underlying agent separately.
Atticus — talk to your agent instead of typing
Hold-to-talk is a small design decision that changes how an assistant feels to use: you press a key, speak, and release. Atticus applies that pattern to macOS, running on top of Claude command-line mode. The transcribed request is passed to the assistant, which makes it a hands-free way to drive a coding or writing session. It is built for people who would rather talk than type when giving an agent instructions. The app is open source and built specifically for macOS, but it does need your own Claude setup to function, and there is no Windows or Linux build.
PDF to MP3 AI Connect — documents you can listen to
Long documents are hard to get through at a desk and impossible during a commute. PDF to MP3 AI Connect converts documents into spoken audio using the text-to-speech access that comes with an existing OpenAI subscription, rather than charging separately for narration. You connect your account and turn PDFs into MP3 files you can play anywhere. It targets readers who want to work through reports, papers or ebooks without sitting down to read them. The obvious catch is that it requires an active OpenAI subscription, and voice quality depends on what that subscription exposes.
Big Arrow on the Screen — let the agent point
Text is a poor way to say "look at this part of the screen". Big Arrow on the Screen lets AI agents draw large arrows, boxes and text directly on top of your display, so an agent can indicate exactly the region it means instead of describing it in words. That makes supervised and shared-control workflows far easier to follow, and it is genuinely useful when you are watching a computer-use agent work. It is open source and easy to drop into an existing setup. The trade-off is that overlays can obscure the very content you are trying to look at.
CUA-S1 — a model built to drive a desktop
Most models answer in text. CUA-S1 is built for computer use: it is trained to operate a desktop interface the way a person would, moving the pointer, clicking and completing multi-step tasks inside real applications. It ships alongside the CUA project, which provides tooling for running and evaluating these agents. The focus is general control of a computer rather than single replies, which makes it interesting for automation that has to work through a normal graphical interface. Running it needs capable hardware, and it still warrants supervision before you let it act on real files.
Pulse Local — budgeting without uploading your bank data
Sending financial history to a hosted service is a reasonable thing to be uneasy about. Pulse Local is a Mac budgeting app whose AI models run on your own machine, so spending data never leaves your computer. It is aimed at people who want AI-assisted budget review without uploading bank or card data to a third party. On-device models are smaller than hosted alternatives, so the analysis is more modest, and the app is Mac only with no mobile version.
What this wave has in common
Two themes run through these launches. The first is local execution: Janus, Engrim, Pulse Local and CUA-S1 all keep model weights or personal data on your own hardware instead of sending it to a hosted API. The second is agent infrastructure. Engrim, Offrun, Big Arrow on the Screen and CUA-S1 are not agents themselves — they are the memory, coordination, feedback and control layers that make agents usable once you move past a single chat window. That shift from an assistant that answers to a set of parts around an agent that acts is the clearest signal in this batch.
How to pick one
Match the tool to the problem you actually have. If local inference across mixed GPUs is the pain, start with Janus. If your agents forget everything between sessions, Engrim is the direct fix. If you are juggling several agents at once, Offrun gives you one place to watch them. For hands-free input on a Mac, Atticus. For listening to documents during a commute, PDF to MP3 AI Connect. If you are building or supervising a computer-use agent, Big Arrow on the Screen and CUA-S1 pair naturally, one showing you what the agent means and the other doing the driving.
FAQ
Are these tools production ready?
They range from early open-source releases to hosted products. The open-source ones expect you to be comfortable with command-line setup, and the hosted ones are new enough that limits and pricing may still be moving.
Do any of them cost money?
Janus, Engrim, Atticus, Big Arrow on the Screen and CUA-S1 are published as open-source repositories and are free to run. Offrun, PDF to MP3 AI Connect and Pulse Local are hosted or commercial products, and PDF to MP3 AI Connect works through an existing OpenAI subscription.
Why source these from Hacker News?
Show HN posts are where a lot of small AI tooling surfaces first, and the vote count is a rough signal of whether other practitioners found it useful. It is a filter rather than a guarantee, so each tool here was checked to confirm it is a real product with a working link.
