Skip to content
Vytherix.Talk to us

Private AI for repeatable work

Everyday work, finished with care.

Vytherix is a calm AI coworker for small businesses that write the same kinds of documents over and over. Give it an outcome, and it turns your notes, files, and repeatable instructions into a clear plan and a finished draft your team can review, edit, and export.

The edits your team makes are the signal. Vytherix learns the structure and tone of your work, tests improvements against examples it did not train on, and rolls back when a change does not earn its place.

See how it worksWork with Vytherix AI Labs
ImprovedVisit notes

Got better at Visit notes

Learned from the drafts your team reviewed. On examples it had never seen, drafts needed less editing than before. Nothing left this computer.

Undo thisOne button. Always available.
An illustration of the improvement card inside the app. Real cards show the workflow, the date, and the measured change.

Built for repeatable work. Supported by the lab when the workflow is specific.

Vytherix is an open-source desktop app for everyday work. A small business can use the generic workflow pack, create its own workflows, and keep the whole improvement loop on one Mac. When documents need a more exact pack, evaluation method, or model choice, that is where the lab comes in.

Run it yourself

The Vytherix app, open source

Point it at the paperwork your team already writes and it drafts from day one, then improves from the drafts your people correct. No terminal, no keys, no account. Nothing you write leaves the Mac.

The desktop app is in active development. Availability and installer access are still being prepared, so there is no pretend download button here.

How the app improves itself

Hire the lab

Vytherix AI Labs, for your specific documents

Deep fine-tuning on your own reviewed work, a pack built around your forms and checks, evaluations that define what “good” means for your business, and a straight answer on the hardware you need.

What an engagement involves

What it does

Everyday work, with a human in the loop

A useful first draft, a clear review step, and a record of what happened.

Built for the work between the real work: visit summaries, customer replies, quotes, handoffs, notices, and follow-ups. It helps with the job in front of you, then gets more consistent from the corrections your team already makes.

01

Workflows, not a blank chat box

Start from a guided job: summarize notes, draft a reply, or extract fields into a table. Add your own workflow when the shipped ones are not enough.

02

Draft, edit, export

Paste notes or choose a local text file, get a draft, make the edits that matter, then copy it or save it as a PDF. The person stays responsible for the final version.

03

A quiet improvement loop

The difference between the first draft and the version you save is the useful signal. Vytherix keeps that loop on the Mac and only promotes changes that hold up on held-out examples.

04

Inbox for decisions

Downloads, updates, questions, and anything the system cannot decide alone wait in one place. Nothing important happens silently in the background.

05

Local schedules

Prepare a recurring draft daily, weekly, or monthly. The schedule runs on your Mac and leaves the result waiting for review rather than sending it for you.

06

A record you can verify

Runs, model changes, improvements, rollbacks, and approvals are written to a local hash-linked record that can be checked and exported for review.

The Mac app

The Mac app

A calm place to turn an outcome into useful work.

Workflows on the left. The work in the middle. A clear view of what the Mac is doing on the right. Vytherix is designed for the person who has to get useful work out, not for someone who wants to manage an AI system all day.

Guided workflows

Start from the job, not a blank prompt.

Review before use

Edit, check, then copy or save.

Visible decisions

Approvals and changes wait in Inbox and Record.

Vytherix Mac app in Codex Dark theme showing guided workflows, projects, Inbox, connections, activity record, and the local Mac status rail.
The current Vytherix Home screen: choose a workflow, review what’s waiting, and see the local engine status — all without leaving this Mac.
How it works

How it works

The draft gets better because someone reviews it.

The product is built around a simple loop: give it the real input, review the draft, save the useful version, and let the system learn only when the evidence is strong enough.

  1. Day one

    It starts with good instructions

    You pick a workflow — a visit summary, a supplier email, a quote. Vytherix ships with instructions written for that job, so the first draft is useful before it has learned anything about you.

  2. After a handful of drafts

    It follows your own examples

    Once a few of your approved drafts exist, Vytherix puts the closest ones in front of the model as the pattern to match. This is where most of “sound like us” comes from, and it happens in the first week.

  3. Once there is enough

    It tests better instructions

    Vytherix writes variations of its own instructions, scores them on your held-out examples, and keeps the one that wins. No training, no waiting, no one asked it to.

  4. When there is really enough

    It trains on your corrected work

    With a few hundred reviewed drafts for a workflow, Vytherix trains a small adapter on your own approved pairs — on your machine, from your material.

The exam comes first

It has to prove it, on work it never studied

A slice of your reviewed drafts is set aside and frozen. Vytherix never trains on it and never tunes instructions against it. It is the only thing allowed to approve a change, and the gain has to be bigger than the noise — not a margin that could be luck.

If it gets worse

The old version comes back by itself

A promoted version serves real work while the previous one stays ready. If people start editing more again, Vytherix reverts on its own and says so. Every previous version is kept, so undo is one button and one plain line of explanation.

Never at the cost of your machine

It waits for a quiet moment

Training only starts when the Mac is idle and on power, with enough disk and headroom — and it stops the moment someone starts working, picking up later. It is not a nightly job that fights you for your computer.

And when there isn’t enough yet, it says so. Below a certain number of reviewed drafts, no test can tell a real improvement from luck. So Vytherix does not train and does not promote — it tells you it is still learning and waits. There is an “improve now” button, but it cannot click past the statistics.

Why local

Why local

Private by default, visible when a decision matters.

Local inference is only part of the promise. The app also makes its boundaries visible: what it read, what it changed, what it needs from you, and what stayed on this Mac.

  • It runs on hardware you already own

    A Mac mini, a Mac Studio, or the Mac on the desk. Open-weight models run on that machine. There is no per-seat meter or usage bill for local drafting.

  • Local by default, choices stay explicit

    Local drafts stay on that Mac. If you later choose a supported provider or connect Gmail, that route is visible, opt-in, scoped, and recorded before it can read or send anything.

  • A person stays in the loop

    Every result is labelled a draft. Vytherix does not send, file, or post it on its own, and the correction that makes the draft useful is also the signal that helps the next one.

  • You can read the record of everything it did

    Every run, every lesson captured, every version that was tried and kept or thrown away, in a searchable list written in plain words. Not a log file — a record a non-technical owner can actually read.

The machine

Where it runs

Start on a Mac. Plan the next machine around the work.

The Mac path is the one available today. NVIDIA support is a future path, so the example on the right is clearly marked as illustrative.

Front view of a compact silver Mac Studio desktop computer.

Supported today

Mac Studio: notes into a family update

After a home-care visit, a coordinator drops in a few rough notes. Vytherix turns them into a structured family update, the coordinator fixes the details, and the next draft follows the team’s headings and tone. The result is still a draft for a person to review.

Vytherix checks memory during setup and recommends a model that fits. 16 GB drafts; 32 GB and up also trains.

Oblique view of a compact NVIDIA DGX Spark desktop AI workstation.

Future path · illustrative

NVIDIA DGX Spark: a larger batch of examples

If NVIDIA support ships, a concrete use case would be testing a larger set of reviewed quote drafts in one batch: compare a few model sizes, check the output shape, and keep the version that needs less editing. Every output would still be a draft for a person to review.

A DGX Spark is not supported today. Do not buy one for Vytherix based on this image; support and the right workload still need to be measured.

The product photos are used under Creative Commons licenses: Mac Studio photo by Yasu (CC BY-SA 3.0), and DGX Spark photo by Daniel Lu (CC BY-SA 4.0). Vytherix is not affiliated with Apple or NVIDIA.

Use cases

Use cases

Businesses that write the same kinds of documents, over and over.

Vytherix is for the work between the real work: the notes, replies, quotes, handoffs, notices, and follow-ups that repeat every week. A person still reviews what goes out — that review is what makes the next draft more useful. Here are a few concrete examples.

Any small business

The daily work between the real work

Whether you run a studio, clinic, shop, service company, or professional practice, every business has recurring notes, messages, handoffs, and follow-ups that still need a person’s judgment.

The paperwork

  • Daily notes turned into clear priorities for the team
  • Customer messages turned into replies for review
  • Meeting notes turned into decisions and next steps
  • Invoices, forms, and documents reduced to fields you can reuse

What it learns: The structure, terminology, and level of detail your team keeps when they review drafts — without forcing your business into an industry template.

Home care agency

The writing that happens after the visit

Carers finish a call and then write it up. Coordinators turn those notes into summaries, care plans, and updates for families — the same handful of documents, several times a day, for years.

The paperwork

  • Visit notes turned into a finished, readable summary
  • Care plan updates written in the agency's own structure
  • Weekly family updates that sound like a person, not a template
  • Handover notes between shifts, in a consistent format

What it learns: It learns how your coordinators actually phrase things — the headings you use, the order you put them in, the tone you use with a family versus a district nurse.

Food truck / market stall

One person, and all of it is admin at 11pm

The cooking is the easy part. The rest is emails to suppliers, quotes for a wedding, an application for a pitch, and a post about where you are parked tomorrow — written after service, when you are exhausted.

The paperwork

  • Supplier orders and the chasing email when the delivery is short
  • Tomorrow's prep list and the stock count from tonight's numbers
  • Catering and event enquiries turned into a real quote, with your terms
  • Pitch, market, and permit applications that ask the same questions every time
  • The week's rota, and the message that goes out when someone swaps a shift
  • The daily post: today's menu, today's location, today's sold-out item

What it learns: After a few weeks of you fixing its drafts, the quotes come out with your pricing structure and your cancellation terms already in the right place, and the daily post sounds like your account instead of a brand.

Trades business

Quotes and job reports, written from site notes

A plumbing, electrical, or roofing outfit wins work on how quickly a clear quote lands, and loses evenings writing them from a photo and four lines of notes typed in a van.

The paperwork

  • Site notes turned into an itemised quote with your standard wording
  • Job completion reports for the customer and for the file
  • The awkward email explaining extra work found once the wall was open
  • Handover and warranty notes at the end of a job

What it learns: Your exclusions, your payment terms, your way of describing a fix without alarming the customer — captured from the quotes you already sent, not from a generic template.

Small property manager

Notices that have to be clear and consistent

A few dozen units generate a steady stream of letters, each of which has to be firm, clear, and worded the same way it was last time.

The paperwork

  • Maintenance and access notices to a tenant or a whole block
  • Contractor instructions written from a reported fault
  • Rent reminders in the right tone for the right stage
  • Move-in and end-of-tenancy letters, and the inspection write-up

What it learns: The structure your letters use, the standing paragraphs you always include, and the difference between a first reminder and a fourth one.

The app ships a general document pack, and for a lot of businesses that plus their own reviewed drafts is enough. When your paperwork is more specific than this — a form with required fields, a report an insurer or a funder expects in a fixed shape, a document nobody outside your trade has ever seen — building that pack and fine-tuning for it is what Vytherix AI Labs does as an engagement.

Independent consulting

Independent consulting

Practical AI advice for businesses that need a useful answer, not a sales funnel.

Vytherix AI Labs also works independently with small teams on local AI, document workflows, evaluation, and model choices. The goal is simple: leave you with a clearer process and a setup that earns its place in the work.

01

Find the right first workflow

Map the repeatable paperwork that costs your team the most time and choose a useful, measurable starting point.

02

Choose the right local setup

Compare models, memory, and hardware against your examples—not a leaderboard or a sales pitch.

03

Build the workflow pack

Shape the inputs, instructions, output structure, review checks, and exports around the way your business already works.

04

Measure and hand it over

Set up held-out examples, show what improved, and leave your team with a workflow they can run and understand.

Need an independent view of where AI fits?

Bring one recurring document, a few real examples, and the Mac you already have.

Start a conversation
The lab

Vytherix AI Labs

When you want it tuned exactly to your work

The app improves on its own from the drafts your team corrects. An engagement is for when you want that taken further — a pack built around your own documents, and the judgement calls that need a person.

  1. We start from work you have already done

    Your reviewed drafts — the ones a person actually corrected and sent — are the material. We go through them with you, work out what separates a draft you had to rewrite from one you signed off untouched, and turn that into a training set. Where there is not enough of it yet, building and augmenting that set is part of the engagement rather than something you have to solve first.

  2. We build or tune a pack for your documents

    A pack is the app's unit of work: the instructions, the fields, the structure a document has to come out in, and the checks it has to pass. The shipped pack is general. Yours is built around your forms, your headings, your required fields, and the things that make a document wrong in your trade specifically.

  3. We measure against your own held-out examples

    Before anything is handed over, we set aside real examples from your business that the training never touched, and score against them: how much editing a draft still needs, whether structured output stays valid, whether it got slower. You get the numbers and the method, not an assurance. If a change does not beat the version you already had, it does not ship.

  4. We tell you what hardware you actually need

    Which open-weight model, at which size and quantisation, on which Mac — measured against your examples rather than a public leaderboard. That includes the unprofitable answer: quite often the machine already on the desk is enough, and we will say so.

When an engagement is worth it

It comes down to four things: the document your team writes most often, roughly how many go out in a week, whether a person reviews every one before it is sent, and what Mac you already have. Where there is a steady volume of one kind of document and someone correcting it each time, there is something for the lab to work with.

Where there is not, the honest answer is that the app on its own will probably be enough, once it is released.

What we will not do

  • Quote a price before we have seen what you write.
  • Promise a percentage improvement in advance of measuring one.
  • Sell you a bigger machine than your examples justify.
  • Take an engagement where fine-tuning is the wrong tool for the problem.
Honest limits

The honest version

Useful because it knows its limits.

We would rather be clear now than surprise you later. Here are the boundaries before you spend time evaluating it.

  • macOS first, on Apple Silicon

    This version targets Macs with Apple Silicon. Other platforms are not shipping yet, and we would rather say that than let you find out after downloading.

  • Drafting and training need different machines

    A 16 GB Mac runs drafts perfectly well. Training needs more memory than that, so on a smaller machine the app simply will not offer it — it improves using your examples and its instructions instead, and tells you why.

  • It learns form, not facts

    Fine-tuning teaches wording, structure, and format — how your business writes. It does not teach the model new facts. If you need it to know what is in your policy manual, that is a matter of putting the right document in front of it, which is a different feature and a different promise.

  • It drafts. A person still sends

    Every output is labelled a draft, and nothing is sent, filed, or posted on its own. Optional connections can provide context only after you approve them; sending still needs its own visible decision.