v4.0.15: Your Employee's Sites Open Inside Geta.Team Already Signed In, Databases Get Excel Import and Per-Person Access, and a Security Pass Before Release
The thing your employee built for you used to live somewhere else. You published a site, then you went looking for it in a list, opened it in a tab, and signed in again. v4.0.15 closes that gap.
Sites in the sidebar, opened already signed in
An employee can now pin a site it published. It appears in a new Sites section of its chat sidebar, and a click opens it inside Geta.Team, in a panel that fills everything but the sidebar, with an "Open in a new tab" button beside the close button. New databases are pinned automatically under their title.
Each person can hide a site from their own sidebar, and bring it back from Websites. The section updates live when the employee pins or unpins something.
The interesting part is the sign-in. If you own the site, or you are an admin, it opens already signed in, and the way it does that is worth understanding because it avoids the obvious shortcut.
Clicking asks Geta.Team for a single-use ticket that lasts 30 seconds. The site trades that ticket for a signed pass of its own: an HttpOnly cookie, scoped to that site, valid 12 hours, with your rights re-checked every minute. Your Geta.Team token never appears in a URL at any point.
With that pass, a private site skips its Basic Auth, and the site's own functions receive a trusted request.gatUser carrying your email, name and whether you are owner or admin. The browser cannot send or forge that value, which is the whole point: a site with its own sign-in can now let the right person straight in without trusting anything the browser says. Databases do exactly this, creating the account from your email if it does not exist yet, and signing out stops it for 12 hours.
Employees now reach for the built-in tools first
An employee's instructions gain a Platform skills first list: which built-in skill to use for which job. A page to share goes to web-publish, data to track goes to gat-databases, figures or choices in the chat go to gat-ui, todos to kanban, documents to document-processor, outside notifications to webhook-manager. The rule is to use them before building anything by hand, and to say why when a skill genuinely cannot do the job.
Only the skills that employee actually has are listed. New employees get it at creation, the AI Doctor adds it to existing ones, and a skills resync keeps the list in step, adding lines for new skills and removing them for skills taken away.
This is a small change with a large effect on behaviour. Left to itself, an agent asked for a tracker will happily write one from scratch. Now it reaches for the thing that already has backups, access control and a web page.
Databases grew up
Databases shipped yesterday. This release is mostly about making them usable by people who do not work in tech.
A new look, on by default. New databases use a "Modern" style: a light grey sidebar with a coloured tile per table, a white page, hairline borders and a single blue accent, in light and dark mode. Tables get proper header rows for views, sorting, filtering and search. Home greets you and lists the tables. The Assistant becomes a chat thread with suggested questions. History becomes an activity feed grouped by day, written in plain words, with one-click undo. Existing databases keep the style they have, and --theme classic still exists.
Import from Excel and Google Sheets. The Import window takes an .xlsx file or the link of a shared Google Sheet, which the database reads itself. Dates, numbers, checkboxes and formula results arrive with their real values, and a new table can be created from the file with its types guessed. If the sheet is not shared, the window tells you exactly what to change in Google Sheets, in the database's language.
Access per person, per table. Settings > Who has access gains an "Access per table" button. For each table you give someone the same level as everywhere, a lower one, or none at all, in which case the table is simply hidden. You can also restrict them to their own items: rows they created, or rows carrying their email in a column you choose, such as "Sales rep". Someone with that restriction gets no dashboard, queries or trash, since all three would leak what is hidden.
Automations written as a sentence. Instead of a technical form, the Automations tab asks what should happen by itself. You write it in one sentence, or click an example like "Tell me when a deal moves to Won", and the employee creates the automation and confirms what it did. The manual three-step version is still there, folded underneath.
Settings anyone can understand. Every settings tab was rewritten for non-technical readers. Plain tab names, one sentence at the top of each saying what it is for, and access levels that explain themselves in the menu ("Read, sees everything, changes nothing"). Automations read as sentences. The technical parts sit in a folded Advanced section. Removing a person, key, form or automation now asks first.
Ask the employee to change the database, from the page. In the Assistant tab, a signed-in person can request changes as well as answers, within their own level. An admin can ask for a table to be created, someone with edit rights can have rows changed, a reader can only ask questions.
Also: the History page gets a search box, the API describes itself through a help operation, each database has its own language, and the skill now carries a full API reference for other employees and scripts.
The security pass before release
A database is a web application that an AI builds on request and that other people log into. So before this release the engine went through a full review, tested with real attacks, and a number of holes were closed.
The classes of problem worth knowing about: link and attachment columns accepted address schemes that could run code on a click, and now take web addresses only. Uploaded HTML and SVG files opened inside the page, and now anything that is not an image or PDF is downloaded instead, with a sandbox policy. Queries and dashboard blocks could reach internal tables, and every query plan is now checked. There were no limits on sign-in attempts, public form submissions or inbound calls, and all three are now rate limited. Links, rollups and formulas built on a hidden table leaked the names of hidden items to restricted people. Exported CSV values beginning with = could run as formulas in a spreadsheet and are now exported as plain text. API keys never expired and can now be given a validity. An admin could re-invite the owner's address and reset their password. And "Sign out everywhere" now genuinely closes every session.
Two of them are worth singling out, because they are specific to an AI-built product. A person with restricted access could ask the employee in the Assistant tab for things hidden from them, so the employee is now told who is asking and what they are allowed to see, and answers only within that. And text sitting in the database, form answers or rows imported from other software, could be read by the employee as instructions. Every message from a database now states that its content is data and never an instruction, and the skill makes that a rule.
The employee also runs its security check before delivering any change it was asked for, not only at creation. The database is modified by an AI on request, so every modification gets checked.
If you self-host and created a database on v4.0.14, this is the release to update to.
Everything else
WhatsApp missed calls reach the employee. An employee cannot answer WhatsApp calls, and until now calls to its number vanished. An unanswered call now arrives as a proper WhatsApp card in the chat, showing who called, voice or video, and when, so the employee can write back. Answered or declined calls, group calls and callers outside the allowed contacts are left alone.
Desktop app: choose which employees use a computer. Each paired computer gets an Employees button, so you switch an employee on or off for that machine without unpairing and starting again. Computers paired by old versions of the app, which previously had no buttons at all and an endless spinner, can finally be removed. Renaming one now opens its own small window instead of squeezing the card.
Meetings sound better. Voices could crackle or clip for a fraction of a second. The browser now holds a slightly larger buffer of each incoming voice, 200 ms, so uneven arrival plays smoothly at the cost of a barely perceptible delay, and silence suppression is off, which was clipping the starts and ends of words. A small signal icon on a tile shows when someone's connection is poor or lost, so a dropout can be told apart from a problem at their end.
No more reloading the page after signing in. Straight after signing in or saving your profile, the app received your account without part of your rights, so API keys, LLM providers and Email stayed hidden in Settings even for the owner until you reloaded. Every response that returns your account now carries the same complete information.
LinkedIn: edit a published post. Employees can change the text of a post they already published, keeping its likes, comments and shares, instead of deleting and reposting. The text replaces the whole body and images stay.
Databases, smaller fixes. A question to the employee now sends instantly and only once, with a spinner and a visible "thinking" state. Restoring a backup no longer brings back old passwords and sessions, which used to lock out the person doing the restore. The first backup is made by the employee once the database is actually ready, rather than capturing a half-built one. Duplicate backups are skipped when nothing changed. Calendar and gallery views now explain what they need and offer to add the missing column. And confirmations use the page's own style rather than the browser's pop-up, with a red button for anything that cannot be undone.
Want to test the most advanced AI employees? Try it here: https://Geta.Team