Getting Started

Portal Agent is a company portal that an AI agent assembles and maintains. You describe the task in words, the agent creates and edits the pages, and the portal itself stores them as plain files under version control. People do not edit content in a browser: there is no editor in the interface — which rules out accidental edits and broken layout.

What Portal Agent is

An ordinary website page is made by hand: a person opens an editor, types the text, arranges the pictures. Here it is different. The page is assembled end to end by a managing agent — a program with artificial intelligence that you tell what should be on the page. The portal stores the result itself: every page is a folder with a description file and separate text files.

All the portal’s content lives as plain files in a version store — the name for a program that remembers the history of file changes (git, for example). Two things important for business follow from this. Every edit has an author and a time, so you can always see what changed and when. And the whole portal can be moved to another server by simply copying the folder — no export, no import, no “database migration”.

What the portal consists of

  • The page tree — a list of all pages on the left, the open page on the right. This is the portal’s main screen.
  • Blocks — what a page is assembled from: text, images, tables, cards, carousels, buttons. 28 types in total.
  • Python scripts — small programs that compute numbers and pull data from your systems. Python is the programming language such calculation scripts are written in.
  • Data and visualisations — if you have a database, a summary on a page can be built from it, with no program: a single query description is enough. The result can be shown in 33 ways, from one number and a gauge to a map of regions (see the Data Sources & Visualisations section).
  • The public site — some of the pages can be opened to the outside at /p/: anyone will see them, without signing in to the portal.

Who it fits

  • Internal knowledge bases — instructions, onboarding, answers to frequent staff questions, kept up to date by the agent.
  • Reference pages and regulations — rules, contracts, document templates in one place and in one consistent shape.
  • Pages with live numbers — warehouse stock, service status, sales summaries: the data is pulled by a script, not typed in by hand.
  • Public sections for clients — news, a service catalog, landing pages, open to the outside at /p/.

First sign-in

Open the portal’s address in a browser. While there is not a single user in the database, the registration form shows by itself — the first person to register becomes the administrator and gets full access, drafts included. The second way is to set the login and password in advance through environment variables at server startup; the variable stores not the password itself but its fingerprint (a lowercase SHA-256 hash). The sign-in session lives for 90 days, so you will not have to sign in every week. An empty portal opens with a “Welcome” greeting and a hint to ask the agent to add pages.

What you see on screen

Almost the whole portal screen is two parts: the page list on the left and the open page on the right. The rest — search, history lists and service buttons — sits on top:

  • Page list — the portal tree on the left, with full-text search and a lists switch (recently viewed, updated, created). Its width can be changed, and the list itself can be collapsed.
  • Open page — the content of the selected page on the right. This is where most time is spent.
  • Comments on a page — under the content there can be a thread of comments: the administrator selects text on the page and notes what to fix, and the agent works through those remarks. The same selection can also send the agent a task as a separate conversation. Visible to the administrator inside the portal only (see the Page Comments section).
  • Chat panel — slides out from the side by the New Chat button in the header: the conversation with the AI agent that edits the portal. Appears only when Xedant Agent is connected.
  • Settings — the gear in the header: appearance (theme, language, font size) and sign-out. Settings are stored separately for each user.
  • License — the license state is shown as a line at the bottom of Settings. It is needed only for sending messages to the agent; everything else works without it.
  • The Commit and Undo buttons — appear in the header when an agent is connected and its working folder holds uncommitted changes. Commit writes the agent’s work into the version history; Undo asks first and then irreversibly reverts every uncommitted change.

The light theme is on by default. The theme and sound toggles, settings and a new conversation all sit on buttons in the header. The same answer from the agent is never played twice, even when several portal tabs are open.

The first steps

  1. Connect Xedant Agent — give its address and access key in the portal’s settings. Without this, the agent chat is not shown in the interface.
  2. Ask the agent to create the first page — describe in words what should be on it. The agent will assemble the page from blocks.
  3. Assign access — a new page is always created as a draft. Until you say otherwise, only you see it.
  4. Mark the public pages — the ones that should be visible from the outside at /p/.
  5. Check the most important things with your own eyes — style, block order and live numbers are the agent’s job, but nothing replaces the owner’s first look.

What the interface does not have

The portal has no page editor: neither visual nor text-based. There is an Edit with agent button at the page header and on every block, but it does not edit text: it opens the chat panel and fills in which page or block is meant, while you phrase the instruction in words. This is deliberate — the content is edited by one managing agent through a separate service address, not by people in a browser. Because of that, pages always stay in one style, and the edit history stays clean: every change is a separate entry in the file history, not a stream of random clicks.

Next: how the tree works and what the reader sees — in the Pages & Tree section; what a page is assembled from — in Blocks; how the agent edits content — in Agent Management.

← Back to the documentation index