Not every AI launch this year is about agents writing code. A second group is quietly about the opposite thing: keeping the work, and the data, on your own computer. Some of these tools run models locally because sending the data anywhere is not acceptable. Others run locally because it is faster than waiting on an API. A few are not local at all but exist to give you back judgement that using AI every day tends to erode.
Eight tools launched between May and October 2026 make up this roundup. They are new to the directory and they share a bias toward the individual over the team.
Before you buy the hardware
Local inference has one persistent problem: you cannot tell how fast a model will feel until after you have bought the machine. Llmspeed addresses exactly that. You browse by hardware — an RTX 5090, a Mac Studio M5 Ultra, a Strix Halo laptop, a DGX Spark — and see the speed you would actually get. It is aimed at anyone weighing a GPU or memory upgrade who does not want to guess from abstract benchmark tables. If you have ever bought a card and then discovered the model you wanted still did not fit comfortably, this is the tool that would have told you first.
Meetings that never leave your Mac
Trace is a Mac app that records and transcribes meetings entirely on your own machine. It captures your microphone, system audio and optional screenshots, then produces a markdown transcript with a summary and flagged moments inline. Nothing leaves the Mac, which is the entire point: it is built for people who cannot paste a client call or a medical appointment into a cloud service. It requires macOS 14.4 or later and Apple silicon.
Keeping your own judgement sharp
There is a specific skill that atrophies when you direct agents instead of writing code, and Fata is built to stop that. It is a spaced-repetition app that drills full-stack fundamentals — frontend, backend, and the technical design that holds up in production — so you can still judge whether what an agent produced is sound. The premise is blunt: you cannot direct what you do not understand. There is a daily ten-question programming quiz, and it is free on the App Store, Google Play and the web.
Spreadsheets both you and an agent can work in
Most real business work still lives in spreadsheets, and agents are bad at them because they usually have to drive Excel through screenshots. Nobie is an Excel-compatible runtime where formulas and cells are first-class to the agent. It is pitched as a runtime rather than a plugin, so the person and the agent operate on the same model instead of one watching the other's screen.
Learning languages from what people actually say
Two tools in this batch attack language learning from the material side rather than the curriculum side. LingoChunk turns native-language audio into study material: it extracts the words from a real recording and turns them into flashcards and shadowing practice. The tool grew out of the author's own study of German and Greek, and it exists for learners who would rather study what people actually say than textbook sentences.
SubSmith does the same for video. You drop in any file — MKV, MP4 or MP3 — and get accurately synced subtitles you never have to retime, then use them to study vocabulary in context. It is for people who would rather learn from films, series and YouTube than from graded course material, and there is a seven-day trial that does not ask for a card.
Design generation without the slop
Diffui generates interface designs from a prompt using a diffusion model rather than an LLM. The company claims nine options in about fifty seconds against one page in three to five minutes for LLM-based competitors, and you can upload existing screenshots so it learns your design language and produces new features that match. The pitch is brand-compliant generation at speed, aimed at designers tired of generic output.
Data analysis without uploading the dataset
MLJAR Studio is a desktop AI analyst that answers questions about your data in plain English and returns real Python code, charts and reproducible notebooks. Everything runs locally, so the data never leaves your computer, and it can also run machine learning experiments for you. It comes from the team behind the open-source MLJAR AutoML package, which is a useful signal if you care whether the thing is maintained.
The pattern underneath
What connects these eight is not a technology but a refusal. Each one declines a default that the rest of the market accepted — that meeting audio goes to a vendor, that a dataset goes to a hosted notebook, that you buy hardware and hope, that a design tool produces one generic option slowly. None of them is trying to be a general assistant. They are narrow tools that do one job on your own terms, and that narrowness is the reason they are worth adding to a directory already full of general-purpose chat products.
