36 native Mac apps.
One vertical library.

Finder is optimised for not alarming you. It hides paths, collapses iCloud state into one ambiguous cloud icon, and tells you nothing about why a search matched. MacManager takes the opposite position: every row can show its exact path, provider, volume, owner, permissions and link target, and every search result can show which part of it matched — the filename, the metadata, the extracted text, or the visual similarity.

iCloud residency is the clearest example. Finder gives you a cloud badge. MacManager distinguishes cloud-only, queued, downloading, downloaded, stale, uploading, conflicted and errored as separate states, and puts Download Now and a safe Remove Download next to them. If a file is not really on the machine, you know before you open a 3 GB video on a train.

Disk usage tools are easy; safe disk cleanup is not. SpaceLens maps the volume visually so large directories stop hiding, then puts an undo step between you and deletion.

The scan is local and reads the real filesystem rather than a cached index.

Duplicate detection is easy to get wrong, and the cost of getting it wrong is a deleted photo. Sift groups exact duplicates, near-duplicates, blurred frames and screenshots, and always shows the candidates side by side before you act.

All analysis runs on-device with Apple's vision frameworks. Your library is not uploaded for scoring.

The snapping is what you expect: halves, quarters, thirds, two-thirds, sixths, maximise, almost-maximise and centre, by shortcut or by dragging a window to an edge or corner with a live preview. The shortcuts match Magnet's, so muscle memory transfers, and all of them are remappable.

Desktop Shapes are the part that is genuinely different. Any manageable app window can be docked onto a cube face; the cube holds and restores those windows, and populated cubes can be placed into orbit around Gravity's desktop solar system. Scenes also record every window across every display and restore the arrangement when you redock.

Online converters ask you to upload the file. For anything confidential that is the end of the conversation. Recast performs 589 conversion paths across 58 formats locally.

No account, no queue, no file-size ceiling other than your own storage.

Import PDF, EPUB, DOCX, RTF, HTML, Markdown, plain text or images. Rather than rendering pages, Easy Reader builds a semantic graph with stable block anchors, hierarchy and page mappings — which is why an annotation stays attached to the right paragraph after you change the font size.

Focus Flow adapts the presentation for sustained reading. OCR, speech synthesis and the intelligence features run on-device, so a confidential document stays confidential.

Most AI desktop tools bundle the model and charge a subscription for it. MultiAgentOS separates the two. You connect through six modes — API key, local server, CLI pipe, OAuth, Terminal, or a local GGUF file — and the app supplies the part that is actually hard to build: a supervised harness of browser, desktop, code, terminal, file and screenshot sidecars.

Every tool call is rendered with its arguments and its result. Nothing runs invisibly. Subagents are configured with an explicit goal, a tool allow-list, a recursion limit, a turn budget and a defined result handoff, so a runaway loop is bounded by construction rather than by hope.

The interface is a conventional browser — tabs, address bar, bookmarks, history, downloads, settings. The difference is a bridge panel where a connected model can navigate, click, fill forms, run JavaScript, take screenshots and report back, with every step shown.

The same surface is exposed as JSON-RPC 2.0 over HTTP plus a WebSocket on port 9224, so your own scripts can drive the browser without a model in the loop at all.

Twenty-eight tracks run from beginner foundations through algorithms, Bash, C++ and web fundamentals. Lesson steps are runnable, so correctness is established by execution rather than by a model's opinion.

The tutor connects to an API key or a local model, which means the whole thing works without an account or a subscription and without sending your code anywhere you did not choose.

The usual workflow is a spreadsheet, a document and a deck holding three copies of the same numbers, which drift apart the moment one changes. InQ keeps one `.inq` file and gives it four views: Flow for prose, Data for spreadsheet work, Canvas for freeform layout and Present for slides.

Because the views share a source, a changed figure propagates to every formula, chart and narrative reference that points at it. The deck does not go stale, because it was never a copy.

Built for thinking on a page rather than filing notes into a database. Handwriting, typed pages and PDF markup share one workspace, so annotating a paper and writing about it are not two apps.

Local-first: notes are files on your machine, with no account and no required sync.

Spider diagrams are a well-established memory technique: spatial position and radial relationships give recall something to hold onto that a linear list does not. Most apps treat mind-mapping as a drawing feature. Here it is the primary model — nodes auto-arrange into a clean web instead of drifting wherever you drop them.

Recall mode is the payoff. It blanks the labels and asks you to rebuild the structure from memory, turning the note into a test.

Transcription runs on-device. The notes it produces are structured rather than a wall of text, and each point carries the timestamp it came from, so you can check the claim against the moment in the video that produced it.

That citation link is the whole design. Summaries you cannot verify are worse than no summary.

Transcription runs locally, which matters most for exactly the recordings you would never upload — client calls, medical discussions, internal reviews.

The archive is the other half. A transcript you cannot find again is a transcript you did not make.

Capture a page with the camera; get structured cards rather than a photo. The extraction produces question and answer pairs you can edit before they enter the deck.

Scheduling is spaced repetition, so review timing is driven by your recall performance rather than by the order you happened to create the cards in.

You describe the video; it assembles a cut with captions and pacing. The important part is that the result is an editable project, not a finished export you have to accept or redo.

Everything renders locally.

Most people keep an editor and a streaming app open at once and duplicate their setup in both. Broadcut merges them: the same sources, scenes and assets serve a frame-exact timeline and a live RTMP desk.

Local rendering and local encoding throughout.

Live front-camera capture with looks applied in real time. The whole pipeline is local: frames are processed on the device, and the face landmarks used to place a look are discarded when the session ends.

The app ships with no advertising or tracking SDKs, which in this category is a design decision rather than an oversight.

Skincare advice is mostly assertion. Radiant does three things instead: it measures the appearance of skin from a guided scan, it builds a routine where each recommendation carries its citation, and it re-measures over twelve weeks so the question 'did this do anything' has an answer.

Everything runs on-device — no account, no upload, no server — which matters more here than in most categories, because the input is photographs of your face.

Fine-tuning fails on data quality far more often than on hyperparameters. Alembic handles the unglamorous part: cleaning, deduplicating, normalising and formatting source material into the structures training tools expect.

Every transformation is a reviewable step, so you can explain what happened to the data and undo it.

The important word is deterministic. Tidyset does not guess at your values with a model. It detects whitespace, inconsistent labels, missing values and type problems, proposes a fix, and shows the before-and-after diff for you to accept or reject.

Every accepted action becomes a step in a recipe. The recipe is simultaneously your undo history, your audit trail, and the thing you re-run when the same messy export lands next month.

Forty courses, eighty modules, four hundred lessons. The structure matters less than the fact that 140 of them are editable code labs that execute locally — you are not reading about a retrieval pipeline, you are fixing one that fails.

The tutor is constrained to the curriculum. That is a deliberate limitation: a general chatbot will happily invent an answer, while a grounded one will tell you which lesson covers the gap.

Forty courses covering ingestion, transformation, modelling, orchestration and warehousing, with the same runnable-lab structure: you build the pipeline and watch it fail before you read why it failed.

Everything runs on-device, including the tutor, so a work laptop with no cloud access is still a full learning environment.

Forty courses across design, testing, code review, debugging and delivery. The labs are the point: you refactor real code, write the test that catches the real bug, and review a diff that contains a real defect.

Runs entirely offline with a tutor grounded in the material.

Forty courses moving from Python fundamentals through PyTorch, training loops, evaluation, serving and MLOps. Labs execute locally so the gradient that does not descend is your gradient.

The tutor stays inside the curriculum, and the whole app works with no network.

220 lessons across HTML, CSS, responsive design, JavaScript and DOM interaction, plus 10 capstone starters with real offline assets. The code runs in a sandboxed WKWebView on your machine, so the preview is a real browser engine.

The checker reads the rendered DOM and computed styles. That is the difference between a course that verifies you wrote the expected characters and one that verifies your page actually works.

Syntax is the easy part and every course teaches it. CodeMonkey Pro trains what comes next: reading code you did not write, refactoring without breaking behaviour, spotting the defect in a pull request, debugging from a symptom, writing tests that would actually catch a regression, and reasoning about design, performance and security.

Eight interactive studios grade deterministically. Python 3.13, JavaScriptCore and isolated SQLite runtimes execute on-device, so the exercises run real code rather than simulating it.

Most learn-to-code games interpret a toy language. This one embeds CPython. You write Python, it executes locally, and the resulting action stream is replayed by the character in the world. When your loop is off by one, you watch it be off by one.

300 levels across 10 chapters and 10 puzzle modes, each deterministic and machine-verified so a solution that works, works every time.

Most language apps are excellent for three months and then leave you circling A2 forever. Talkify carries 21 languages through a full A1–C1 path: 30 units, 1,075 sessions, 750 phrase sets and 750 four-turn dialogues per language.

Each session mixes listening, phrase building, meaning recall, pronunciation practice and — where the script requires it — character tracing, with adaptive review scheduling what you got wrong.

Pronunciation apps teach you to say a word correctly. Accent Coach teaches the systematic features that make an accent that accent — the six to nine signature sounds of London English, General American, Cockney, Parisian French or Beijing Mandarin — with IPA notation and physical production guidance for each.

Then you practise: hear the reference at normal or slow speed, record with a countdown, get scored, and A/B your take against the original. Scoring runs on-device, so your recordings stay on the machine.

A pronunciation score of 72% tells you nothing actionable. Mimic renders pitch, rhythm, delivery and articulation as visible traces, so you can see that your intonation falls where the reference rises.

No account, no analytics SDK, and no cloud scoring at runtime — the comparison happens on the device holding the recording.

Each of these languages has a small set of features that reliably defeat learners. Mandarin: four tones plus neutral, tone sandhi, and the q/x/j and zh/ch/sh/r distinctions. Japanese: pure vowels, mora timing, long vowels, the small tsu pause, devoiced vowels and pitch accent. German: ä/ö/ü, both ch-sounds, the uvular r, final devoicing and w/v/z/s.

The format is testing rather than lecturing. You are asked to produce or discriminate, and told whether you did.

Written for people whose career runs through a second language — doctors, nurses, lawyers, sales staff, managers, support agents. You pick a profession and a scenario and hold the conversation aloud against an animated avatar.

Everything runs on-device using Apple's frameworks: recognition, synthesis, and the model driving the exchange. No API keys, no cloud, no per-token bill, and no recording of a clumsy rehearsal sitting on someone else's server.

Roles come from public APIs, ATS boards, company career pages, pasted URLs, browser-assisted extraction or manual import. A candidate profile built from your CV and constraints is used to rank them, so the list is filtered by your actual requirements rather than by keyword soup.

Drafting produces a tailored email and cover letter per role. Nothing leaves the machine until you approve it. Sending goes through your own Gmail, Microsoft Graph or SMTP account with spacing limits, quiet hours and an audit log — so a mass-application script cannot run away with your professional reputation.

Price comparison usually means new retail only, which ignores where the actual saving is. DealZ puts new, refurbished, second-hand and private-sale listings in the same comparison, alongside flights, stays and local services.

Scan a barcode or QR code to resolve the exact product. Crucially, the app explains its provider coverage — which sources it queried and which it cannot reach — so you know the boundary of the comparison instead of assuming it was exhaustive.

Gig platforms insert themselves between the two people doing the work and take a percentage for it. GigaJob is a board: someone posts, someone takes it, and the two of them settle directly.

Native app across Mac, iPad and iPhone.