Changelog

What shipped, and when.

What changed in the platform, newest first — including the changes that make it stricter rather than larger.

Last updated 13 September 2026Subscribe by RSS

Week of 7 September 2026

  • Workspace Start: free for the entire organisation. No seat count, no five-user ceiling — every person in the company can work inside the governed layer for nothing. Leaving colleagues outside it to save money is exactly how shadow AI starts.
  • Workspace, priced per person who actually uses it. Three tiers: Start free, Team at €3.99 and Pro at €9.99 per user / month on annual billing — with shared spaces, team agents and versioned artefacts in Team, and SSO, policy rules and the retained record in Pro.
  • Inspection that fails closed, under any load. When an inbound check cannot finish inside its time budget the tool result is withheld rather than delivered, and the event is counted where you can see it. A policy set to block stays a promise at the moment the system is busiest — which is the moment it is tested.
  • Identity your directory controls, end to end. People exist in the Gateway because an administrator, an invitation or a sign-in through Entra ID, Active Directory or Google Workspace put them there. An application asserting a name nobody knows gets no identity and no budget: it runs as the application it is, and the attempt is counted.
  • Prompt-injection detection 45× stronger, on hardware we own. Inbound content is inspected by a classifier we train and run ourselves. Against BIPIA, the public benchmark for indirect prompt injection, across four languages: 45 times the attacks caught, at zero to one false alarm — and the content never leaves the boundary to be checked.
  • Pricing you can read before you buy. The Gateway is priced by the requests that run through it, never by seats: pick a monthly volume and every step’s price is on the card before you choose it. Nobody is charged as a seat and no application as a connection.
  • The decision about your data is made inside your boundary. Judging whether personal data, financials or confidential business may leave is done by the trained inspector on the content itself, naming which of the nine categories fired — with nothing handed to a third party in order to ask.
  • Proof that your policies were enforced, not just configured. The security view reports how many inspections ran and how many were skipped, by place and reason, over the admin API as well as on screen. The auditor’s hardest question — how often did this control not run? — gets an answer with a number in it.

Week of 24 August 2026

  • A Twin that gets sharper the longer it watches. Knowledge from repeated work is consolidated rather than piled up, so a Twin that sees the same process forty times holds one clear account of it — and every part of that account stays inspectable.
  • An approval anyone can answer in ten seconds. A review step presents the question, what the automation proposes and the evidence behind it, with two buttons — and triggers are configured as three plain questions, because the person who owns a process is rarely the person who enjoys configuring one.
  • The diagram you draw is the screen you watch it run on. The picture of a process and the record of it running are one thing, not two: each step shows what it did and what it cost, in the shape you designed.

Week of 10 August 2026

  • Connect a system in two fields. The Gateway asks the server what it supports and fills in the rest itself, including the discovery paths some providers put where the RFC says rather than where everyone else does. Two systems exposing a tool of the same name no longer collide.

Week of 3 August 2026

  • A complete record of everything the AI did in your systems. Every call through a connected system is readable in the console — which tool ran, for whom, with what result and at what cost. "What did the AI do in our CRM last Tuesday" is a question with a screen that answers it.

Week of 20 July 2026

  • Spending that stops at the limit, per person and per team. Every request is attributed to a person, a team, an application or an agent, so a ceiling set against any of them is enforced as it is reached — not reported once the month is over.
  • A named person signs before anything acts. Tools that write, send or spend sit behind a human signature: the model proposes, a person accepts, and the record keeps who accepted and when. Agents carry their own step budget, so one that goes wrong runs out of room long before it runs out of consequences.

14 days · no payment card · bring your own provider account