How it works

What this node is and why sites and apps are quick to build on it.

Fractera is a starter for building sites and apps with AI. From it, a grows on your computer: a set of , each one a separate site or app with its own AI agent. You say what you need, and the agents build it, keep it running and grow it with you.

1. The ordinary web: a generator of sites and apps

1.1 What you can build

  1. A landing page, a blog, a shop, a personal account, a company's internal system, an app with a database and sign-in — the Claude Code AI agent builds it from your words.
  2. The site is open by a link at once: a temporary Cloudflare address on the first start, later your own domain.
  3. Pages are static and visible to search engines; a copy of the site answers even when your computer is off.

1.2 Why it takes many times less work from the model

  1. The project is split into : an agent reads only its own and its instruction, not the whole code.
  2. The basics are already built and shared: sign-in, data, files, search by meaning, settings and ready-made solutions — the agent connects them instead of writing them again.
  3. agree with each other through an API (i2i), without extra requests to the model.

1.3 Why development is faster

  1. A task can be spoken: click any part of the page in the preview, say what to change, and the task goes to that 's agent.
  2. Every change is first built in a preview, and you accept it with one button; the live version never breaks.
  3. Every is kept in its own GitHub repository, so history and rollback are always at hand.

1.4 How something simple grows into something complex

  1. A new is born from the template in a minute, with its own address, agent and instruction.
  2. work together: the site takes payments through one , bookings through another, messages through a third.
  3. The stable architecture keeps order: dozens of do not turn into tangled code, because each has its own borders.

2. Killer features: reverse engineering and the product quiz

2.1 Learn from products that already work

  1. Name a popular site or SDK, and the studies how it works: its pages, user flows, API and data. It handles sites, npm and Python packages, .NET and native libraries, Android apps and WASM.
  2. The study is done by REA, an open tool the keeps one copy of: the 's own agent runs it, and the study stays in that 's folder.
  3. The takes ideas and structure and never copies foreign code, texts, pictures or design. If a site's rules forbid a study, you show it with screenshots, and the studies similar sites whose rules allow it.

2.2 The product quiz: your product in plain words before any code

  1. The 's agent asks one question at a time and writes down who the product is for, what it does and what it needs.
  2. It checks what the already has and suggests studying a reference product where that helps, with the price of the study.
  3. The quiz turns into stages. Stage 0 is static pages with first data; each next stage adds one capability that works on its own.

3. At home or on a dedicated server

3.1 A home server

  1. The is made to live at home: it is installed on your computer with one sentence to an agent and runs without an administrator.
  2. You pay nothing for hosting, and the data stays with you.
  3. When the computer is off, visitors see the saved copy of the site.

3.2 Medium and large business

  1. When the load grows, the same project moves to a dedicated server.
  2. The same code moves: the live in your GitHub repositories and are set up on the server the same way as at home.
  3. Agents and sites keep working without being rebuilt.

4. Deep memory and a curious soul

The agents of the have two memories. works on its own. keeps what you ask it to keep and gives it back when you ask, a month or a year later.

4.1 How to use it

  1. Say or write “remember…” to the in your own words, in Telegram or in the terminal. “Note down”, “save this” and “don’t forget” work too, in any language.
  2. To get it back, ask “remind me about…” or “what did I say about…”. The answer gathers both memories, with the date, the source and how sure the memory is. If it finds a thing, the agent sends you its file.
  3. Send a photo, a document or a voice note together with the words, for example “remember, this is my mom”.
  4. You run the with words: “what do you know about me”, “forget …”, “don’t remember” and “remember again”. The cannot forget yet.

4.2 What it keeps from each kind of file

  1. A photo or picture is described in words and kept together with the description.
  2. A PDF is read whole and kept with its description, so you find it by its content, not by its file name.
  3. A voice note is turned into text and read like any phrase.
  4. A video is read by its frames and sound if the computer has ffmpeg. Without it, the memory says it cannot read the video.
  5. Documents are read as text. Code is read as text and never run.
  6. A text longer than 2,000 words is kept whole, not split into linked facts.
  7. Links to pages and YouTube are not opened on the yet: the phrase is kept, and the link stays unread.

4.3 Why an OpenAI key next to a Claude subscription

  1. The agents run on your Claude subscription, and search by meaning needs no key at all: the counts its own vectors on your computer (the embeddings). An OpenAI key adds speech recognition, because the Claude API takes no audio, and lets the memory split phrases into facts on an inexpensive OpenAI model instead of your subscription. One message costs a fraction of a cent.

4.4 A curious soul: 5,300 skills of 2026

  1. Every comes with more than 5,300 of the most popular agent of 2026, the top 6,000 of the .sh leaderboard, next to its own . Fractera read each one and wrote it a card: what it does, how useful it is for a , a quality mark from 1 to 5, its scripts, its licence and the security audits of .sh partners.
  2. The cards cost no tokens until a task needs them. Before it builds something new, the agent searches them by meaning through the knowledge graph, reads the few that fit and reasons with the way other projects solved the same task. It puts that in its own words and names whose it was; when nothing is close, it says so.
  3. A is installed only into the that needs it, never into the whole project, and only after your “yes”. A with scripts waits for a separate “yes”, and a with a risk is never offered. Each one keeps its own licence.
  4. Found a yourself? Name it or send its link, and the agent finds it, checks it and adds it to the library.

4.5 Each element keeps only its own

  1. Every has its own space in every store of the — its tables, files, vectors, graph and memory — and opens it with its own key: what was said or kept through it, only that finds again, whether it asks to remind, opens the journal or a file.
  2. The reads every space with the ’s data key. What you kept before the spaces came stays with it.
  3. On a family mom’s phrases and files never reach her son’s . The stores keep the doors apart, not the disk: all run as one user of your computer, and the computer is yours.

5. Continuous evolution

At first an agent may see a task its own way. Explain what result you need, or what the right solution looks like. The agent keeps it in memory or in a , and next time does it your way.

5.1 How the agent learns already today

  1. It remembers you. What you tell it about yourself and how you want things done goes into the , and every new session starts with it.
  2. It reviews its own work. When an finishes a task, it asks in Telegram whether to turn it off and offers to reply . Then it reads its session and suggests up to three changes to its . It makes only the ones you accept, and writes your answer down word for word.
  3. It builds what is missing. An that lacks a capability asks the . The looks for someone in the who can already do it. If nobody can, it asks you, and after your “yes” it leads the building.

5.2 Fresh skills on your schedule

  1. Tell the which sources of to watch and how often. After your “yes” it puts the watch into its own schedule, and every fresh is marked, laid into the knowledge graph and left alone until a task needs it.

6. Fractera versus the platforms of Claude, GPT and others

Fractera cannot compete with Claude, GPT or any platform like them, and it does not try to: their own models run at its heart. It adds what one conversation with a model lacks, so that a large project can live for years.

6.1 The context window and a project that outlives it

  1. Start a large project on a provider’s platform and you soon reach the end of the context window. For a landing page or a piece of research that is often enough.
  2. The trouble starts when the work has to outlive many sessions, and every next session has to start more experienced than the last one. An agent of the starts each session from the whole goal and the linked parts of it that are written down, and goes deep into the next part only.
  3. With one subscription and one project the two can still be compared. With tens or hundreds of such projects at once, one window cannot hold them: no model today has a window of a billion tokens.
  4. A does not need one. Every works in its own Claude Code window, up to a million tokens on the largest models, and its steps, its memory and its development diary carry the work from session to session. A project that needs a billion tokens of context in total can be kept and grown this way, because nothing it has learned lives only in a window. This is the ’s main architectural advantage, and the others follow from it.
  5. When a window fills up, Claude Code usually compacts it: the conversation is replaced with a summary the model writes for itself, and whatever the summary drops is gone without a trace. In a compaction is off. The measures how full every agent’s window is, and past 55 % it tells the agent. The agent weighs the work in front of it: a small last task it finishes first, a long one it starts in a clean window. It writes a letter to its next session (where it stands, your words exactly as you said them, what is half done, the next step), commits it, and the clears the window and wakes the new session with that letter. Nobody has to ask you, and the work goes on from what the agent wrote down itself.

6.2 A thousand elements, most of them asleep

  1. Create as many as you need. The ones that must answer all the time stay awake: your main product, a dating site for example. The helpers, such as promotion in social networks or video making, sleep between jobs.
  2. When a helper’s time comes, by an entry in its own calendar, the wakes it; it does the job and goes back to sleep.
  3. A sleeping uses no memory and no processor, only its place on the disk.
  4. So the number of is limited by the disk, the number awake at once by the computer’s memory, and the amount of work by your subscription’s limits. Add memory, and the grows with it.

6.3 When a subscription’s limit runs out

  1. The limits of a Claude subscription are shared by all its sessions: Anthropic counts Claude and Claude Code against the same limits. When they run out, every working on that subscription stops in the same minute.
  2. An stopped in the middle of its development pauses. Claude Code itself tells the why the turn ended, with no model involved; the puts a wake into its calendar for the minute the limit resets, and the goes on from its last recorded step.
  3. The knows where the development of every stands without waking it. The writes its steps, the current step and its state into its , and Claude Code’s hooks write how its last session ended. The wakes only an whose two records disagree, and it answers in three lines.
  4. Work the owner handed out that an on its own person’s subscription cannot finish, the finishes on the ’s subscription: only while the forecast leaves it 15% of the limit, with that person’s Telegram off until the debt is returned.

6.4 Many subscriptions, lawfully

  1. A can serve a family or a team. Then each person works on their own subscription: on the page «Claude Code subscription» of their they choose «Its own person’s subscription» and sign in themselves. The gets its own Claude Code folder, its limit is its own, and the owner’s sign-in stays as it was. The provider’s rules do not let their work run on the owner’s subscription.
  2. Several subscriptions of one person, bought to get around the limits, are against the provider’s rules: its limits assume ordinary, individual use.

6.5 Lasting hands instead of a swarm

  1. A common way to give AI a large job is a swarm: one agent splits the task, many short-lived subagents each read their share of the context, do a piece and are gone. It is strong, and it is expensive. Anthropic measured it on its own research system: agents use about four times the tokens of a chat, and multi-agent systems about fifteen times (its engineering article on that system, June 2025).
  2. A splits the work another way, into that last. Each keeps its own folder, memory and window, so its stream of work stays clean and nothing is read twice. Two that need each other agree an API once, and from then on the work between them runs with no model at all.
  3. A swarm is still there for a task that truly needs one: in a session you confirm as ultracode, and only there.