An app you install on your own computer, where AI agents do real work for you instead of just answering questions. It remembers how you work, so you explain a thing once rather than every morning.
Where we are heading is one harness a beginner can start using on day one and an advanced user can open up fully. That is the direction we are building toward, not something every part of the Beta does today.
The largest change to the app since the first beta: a rebuilt interface, one account across the site and the app, and Cortex that can write to your vault when you say yes. It is available now as a Beta: public early access. Expect bugs. Keep a backup of your vault.
NeuralVault is a desktop application. You open it, point it at a folder of your own notes, and that folder becomes what it knows. Then you stop typing questions and start handing out jobs: read the notes in these three folders, work out what keeps coming up, write it up and leave it on the board for me. It goes and does that, and comes back with the work and a record of how it got there. The trade calls this an agent harness. What it is for is simpler.
You explain your project, your stack and the way you like things done. Tomorrow you open a new chat and explain all of it again.
It keeps what it learns as plain Markdown on your own disk, so you say a thing once.
Getting a useful answer means pasting your work into someone else’s browser tab.
It works on your own material where it sits, and changes to your files wait for your approval, and plain web searches run without asking.
It answers, and then you still do the work. Every single step goes back through you.
You hand it a job instead of a question. It splits the work, runs to a schedule, and reports back.
When it does act on something, you cannot see what it touched or undo it afterwards.
Every tool call is logged and every approved file write is a checkpoint you can rewind to.
The problem with agents is rarely the model. It is the agent inventing a tool it does not have, forgetting on Tuesday what you told it on Monday, and being impossible to leave alone with anything that matters. Those three failures are what the three below are answers to, and each one comes from how NeuralVault is built rather than from a feature that could be added to anything else next quarter.
Any major cloud provider on your own key · or a local model, if you want one · keys held in Windows Credential Manager, never a plaintext file
Most of the field rents you a model in somebody else’s cloud, and your material has to travel there for anything to happen. This installs on your computer, and the model is your call: a cloud provider on your own key, a model running locally, or one of each depending on the job. Either way its memory is a folder of Markdown files on your own disk.
An agent’s rules are a configuration you can read · audit log and run ledger on every tool call · a snapshot before every change you approve in chat
Every agent works to a rule set: what it is for, which tools it is allowed to use, and which folders it may touch, held in a configuration you can open and read. What it remembers is a note. What it did is a log. Nothing is kept out of sight, which is the thing that makes handing it real work reasonable rather than reckless.
A shared board · a cron scheduler and file triggers · an audit log and run ledger
The harness is the product here rather than a feature added to a code editor. So it is equally at home on research, client files, contracts or a decade of notes, and there is somewhere for the work to live once an agent has done it.
This is the difference between asking an AI something and giving it a job. All four are the harness itself, so all four are true across every edition.
Agents get a defined set of things they can reach, and nothing outside it. You decide what is on that list.
It does not start cold every session. What it learns about how you work is kept, and it gets sharper the longer it runs.
Reading is usually open. Anything that changes a file or runs a command waits for you on a card. Plain web searches run without asking.
Every action is written down. You can go back and see what ran, when, and why it decided to.
Pick one. We will show you what it does for that, and take you to a page made for it.
You want agents that do real work in your repo, on your terms, without surprises.
An agent harness you can read, script and constrain.
An illustration of a screen, with made-up content, not a screen capture.
One email when a NeuralVault version goes out. We note which page you signed up from, so we know who to write for.
A harness is only worth running if you can see what it did and stop it doing more. These are the mechanisms, each stated with its limit.
What ran, whether it was allowed, and what came back are written to an audit log and a run ledger on your disk. The audit log is readable on every edition, Free included.
When Cortex asks to create, append to or move a note, a card shows what it asked for and which files it touches, and nothing happens until you approve. Deletions are refused outright.
An approved write is snapshotted first and one control puts the files back. It restores the files an approved chat write touched, not the conversation: the transcript still shows what was written.
In developmentCovers approved chat writes only. Scheduled agent runs capture a snapshot but have no restore control.The model catalogue shows capability scores beside a per-vendor data-handling label, on every edition, so the choice rests on the facts and not on the brand.
Say “remember that I prefer short answers” and the fact is written to a durable profile that is read back into every later conversation. The write waits on your approval.
A standing rule is appended to your own CORTEX.md, the file Settings shows you, so you can read it, edit it or delete it. It also waits on your approval.
Your notes are plain Markdown on your disk, searched by keyword and by meaning together, and read back into the conversation when they are relevant.
Every action an agent wants to take goes past the guard first. It uses fixed rules, so it gives the same answer every time, and no AI model is asked, so text hidden in a file cannot talk it round.
We improve this every single day. On our own test suite, 8 October 2026. See every test set
You can stop being asked about a kind of change. You cannot stop being asked about the dangerous ones.
A card that allows it offers to stop asking, for that kind of change, in that chat only.
The app says what will now run without asking, and where to undo it.
The Safety guard page lists the steps that wait for you even after you said not to ask again.
A card the guard marks high risk never shows Next time.
Not the category, the software. Every surface below is named rather than described, so you can check it against the product instead of taking the sentence on trust, and anything that does not run yet says so.
It reads what you have written and answers out of it, pointing at the note the answer came from rather than producing something plausible. The search is hybrid, so an exact string and a vague description both land.
Each agent carries its own instructions, tools and folder limits. Give Cortex an objective and it hands focused pieces to sub-agents of its own, each restricted to the tools its role allows, and works over what actually failed rather than stopping at the first error.
Four lanes, from To Do to Done, shared between people and agents. An agent picks up a card and moves it across on its own. Only a person closes one.
Work on a schedule, or triggered by a file changing. It does not need you sitting in front of it for anything you have already approved the shape of.
Every change it asks to make is snapshotted before it happens, and one control puts the files back to how they were before that turn. Being able to undo it is what makes approving it reasonable.
Every product page tells you what a thing does. This is the other half: what is true today, what is not finished, and what happens to your material if you try it and walk away. A teal marker is live now, a purple one is dated below.
NeuralVault is built and tested on Windows. macOS and Linux are experimental: we test them far less, so they may be more unstable. There is no macOS or Linux installer yet.
Public early access, in active development. Expect bugs. Keep a backup of your vault, and do not make this the only copy of anything you cannot lose.
The installer is code-signed by Perpetual Technologies AI Ltd. Windows may still show a "Windows protected your PC" panel while a new certificate builds reputation: choose More info, check the publisher it names, then Run anyway. Nothing is wrong with your machine and nothing is wrong with the file.
Use a cloud provider with your own API key, or run a model locally through Ollama, or mix the two per job. Neither is the default and neither is required. Whichever you choose, the key is yours and Windows holds it rather than a file on disk.
Point it at a folder and it reads and writes that folder in place, as Markdown. No database and no import step, so there is nothing to export if you walk away. The folder was always yours.
NeuralVault sends anonymous app-health and usage counts so we can check every feature works and spot sudden spikes. No conversations, notes, file names or account details, and nothing that identifies you or your machine. A crash writes a log on your own machine and stops there. If you send us a bug report, you see the exact text first and press submit yourself.
Team channels and email have no backend in the app today, so nothing is read, drafted or sent from it. The app sends no message of any kind.
Personal is a flat monthly price, in full on its own page. Business is priced to the size of the deployment and how much of it you switch on, then quoted after a free consultation. Free is the same build with the paid parts locked.
A Markdown vault you can search, chat with Cortex, and the audit log. Running today, in beta, on Free.
The chat-first build is public. One installer, a free account, and the paid rows locked rather than removed.
Most of the harness unlocked: the board and its runner, the scheduler, Orchestra, the dashboard and connectors. Model routing on a usage balance is a separate purchase inside the app, so you can use a cloud model without opening provider accounts and managing keys yourself.
The sync server, per-user accounts, role-scoped agents and the connector timeline, installed on hardware we specify and keep running.
Dates are what we are working to, not a contract. This is a small company shipping a beta, and a date that slips will be changed here rather than quietly left up.
Personal includes 31 capabilities. Business includes all of them and 16 more, 47 in total. The cards list the main capabilities and the totals also count the smaller ones. The bars show each edition as a share of that 47. Business is planned and not yet released. Business testing opens in December 2026, and launch is planned for the first quarter of 2027, with January the goal.
of 47 capabilities
Personal is all your AI bills in one place: we handle the billing, the routing and the backups, and routing is only available through Personal. The NeuralVault AI score and our full model picker come with NeuralVault routing. With your own keys you get a plain model list. You can also use your own keys, such as OpenRouter, OpenAI and Anthropic, and that is free to use with your own keys. Routing makes it easier and nicer. During the Beta everyone has Personal. In the next few weeks, when the Stable version is released, Personal becomes a monthly subscription. We will tell you before it changes.
The complete harness. Nothing here is held back for a higher tier.
of 47 capabilities
Everything in NeuralVault Personal, and then:
The same engine as Personal, pointed at running a company instead of a codebase.
Each page ends with what that edition costs. Personal is a flat monthly price you can see in full; Business is priced to the size of the deployment and how much of it you switch on, and quoted after a free consultation. Free is the same application with the paid capabilities locked, so trying it costs nothing either way.
This is not a lite version and a pro version. Personal is the complete harness with nothing held back: the same agents, the same memory, the same audit trail. Business takes that exact engine and adds what stops being optional the moment other people depend on your output. So the question is not which product is better, it is whether you are running a codebase or running a company.
The full harness, on your machine. Use our model routing and there are no keys to set up; bring your own API key or a local model and there is no per-token bill from us.
Put as much of your own material into it as you like. What it learns is written to your own disk as plain Markdown you can open, edit and back up, not to a row in somebody else's database. Run the model locally too and none of it leaves the machine at all.
Not only for code. Point it at research, client files, contracts or a decade of notes: anything you would rather not paste into a browser tab.
Everything in Personal, unchanged, and then the layer a company needs around it.
The difference is not more features, it is that the work stops being yours alone. Other people can see it, act inside it and be held to what they may touch, and every one of those actions leaves a record. That is the part nobody needs on their own machine and nobody can run a company without.
Both pages end with what that edition costs. Free is on general download now.
The editions page puts Free, Personal and Business in one row with what each one gives you and what it costs, then a section on each. It is the shorter way to place yourself than reading two product pages to find out which one you are.
Compare the three editions