Prax is being built to turn what you describe into working software. Explore how it’s designed to plan, build, test and maintain it.
Scroll · the whole product, chapter by chapterPrax is a permanent engineering team you talk to. It understands what you mean, plans the work, writes the software, tests it in a sealed room, delivers it, and proves what it did. Then it stays: watching, fixing and improving.
Prax is two halves built as one system. Brain is the app you talk to: it remembers your project and tells you when an idea is unwise. The Engine is a governed line of twelve stages that plans, writes, tests and delivers the change. Everything either half does lands in one record. Select any part.
Sarah runs a bakery and has never written code. Her customers phone in orders and she writes them on paper. Follow her request through six moments, in the order they happen. The buttons inside each screen work.
The same six moments happen whether you ask for a bakery system or a change to a payments service. What differs is how much of it you are shown.
A shop owner, an engineer and a compliance lead open the same product. The engine underneath does not change. What changes is how much of it each person is shown, and in whose words. Switch between them. The controls inside the window work.
A translation layer sits between the engine and the reader: 62 concepts, three translations each. Failure is never prettier in plain language. If a check did not pass, every depth is told it did not pass.
In the full app every item in the sidebar opens a working screen: spending, history and undo, the policy simulator, the audit log, the stop control. There is a three-minute tour.
Brain is the half of Prax you talk to. It keeps what it learns about your project as separate memories, each with an ID, a source, a confidence and a track record. Ask for something and it pulls the few that matter, builds the answer from them, and cites every one. Pick a request, then open any memory.
Agreeing is the cheapest way for an AI to look helpful. Brain is built the other way round. When a request is possible but a mistake, the agents that scope, plan and build must say so, with evidence, before they build. Then the decision is yours, and it is recorded.
“Like a world-class architect who listens to what you want, then tells you why the load-bearing wall needs to stay and offers three beautiful alternatives.”The standard every Brain personality is held to
The Brain page has the rest: four memory layers, the evidence check every answer passes before it is spoken, how it tracks what caused what, and the workspace where you steer it mid-flight.
Eight are AI agents: they read, write and review. Four are fixed systems made of plain code: they cannot be persuaded, flattered or prompted. Each stage has one job, one named output and its own signed entries in the record. Walk the line, or pick a stage.
Agents understand, write, review and close. They are good at judgement and they can be wrong, so none of them gets the last word on what is allowed.
Code, not judgement. The same plan gets the same verdict every time. Two of the Policy Engine’s 22 checks call a small AI model, and those two can only add scrutiny. An instruction hidden in a README has nothing to persuade.
The first stage classifies every request as low, medium, high or critical risk. The class picks the lane. The lane decides which stages take part. It never switches off a policy check. Pick a request, then try to send it down a faster lane than it deserves.
The Engine page takes you inside a full 3D hall of the twelve stages, then through the 22 policy checks, the sealed room, the handoff interlocks and the nine layers the stages are one part of.
A system lives for years after it is built. Prax has five modes for that life and a sixth for working inside any of them, side by side. Each one takes something in, runs a known set of stages, and hands back a named record. Pick one.
The Modes page opens each one: a full audit report you can filter, a week of maintenance hour by hour, the router that decides what may fix itself, and a session where you and Prax trade control of a file.
An AI that writes code is no longer rare. Neither is a rule engine, an audit trail, or a choice of AI models. Every one has precedents, and we do not pretend otherwise. What is different is that Prax is designed as one system, with one policy boundary around seven things that are usually sold apart. Select a part to see what leans on it. Then pull them apart.
Most tools govern the agent.
Prax governs the change.
Controls on an agent decide who may run it, what it may touch and what goes into it. Those are useful. They are not the same thing as deciding whether the change that came out is allowed to exist, proving that decision, and looking after the result for years.
In Prax the Policy Engine is stage seven of twelve, in the path of the work, not beside it. Nothing is written to your repository until it has ruled. If the check cannot run, the task is blocked.
The page behind this goes one capability at a time: failures that are diagnosed before they are retried, a rehearsal on a copy of the services you depend on, two AI models asked in parallel on critical changes, if you switch it on, and a plain list of what Prax is not.
Every stage writes timestamped entries to one record and signs each with Ed25519. Each entry’s SHA-256 hash includes the hash of the entry before it. So a record edited later no longer matches its own seal, and nothing after it can be vouched for. The hashes below are computed in your browser. Select an entry, then tamper with it.
Twenty are core controls every change passes. Eight guard integrity and alignment. Thirteen adapt to your stack. Filter by where each one runs, then select any of them.
Stopping makes things safer, so automation is allowed to do it. Resuming re-opens the risk, so no automated process, no AI agent, no scheduled job and no API call from a non-human can do it. And how much Prax does alone is a level you set. It changes who is asked. It never changes what is checked.
Not one red button. Four states, from normal to lockdown, aimed at one task, one run, your workspace, one part of the platform, or everything. When several are set, the most restrictive wins.
Every account starts at Smart. Full is opt-in, earned per repository, and withdrawn automatically the moment one of its conditions fails. Getting it back takes a person.
The Trust page covers the rest, and most of it you can try: the sealed room where AI-written code runs, ten layers against prompt injection, the data commitment clause by clause, your money, and a plain list of what Prax does not claim.
None of these counts customers, speed or lines of code. Each one counts something that was designed on purpose, and each has a reason.
The first screen asks one thing: “What brings you to Prax?” Nobody is asked how technical they are. The answer sets how much detail you start with, and you can change it at any time. Every path gets the same checks and the same default autonomy.
Nothing in the engine knows which code host, which AI model or which hosting provider it is talking to. Each sits behind one neutral interface. Adding or replacing one means writing an adapter. It does not mean changing the engine.
Including the ones where the answer is “not yet” or “no”.
There is nothing to buy today. Email us to ask about early access.
Ask about early access ↓