What we shipped in June 2026
Last month I said I’d stop once a month and tell you what actually happened, in plain language rather than commit-speak. So here’s June — and June was even busier than May: nineteen releases, from v1.8.0 to v1.22.0. If you just want the granular list, it’s in the changelog. Here’s the story behind it.
If May was about filling in features, June was about a shift underneath them: the agent loop stopped being one feature among many and started being the product. And in the background I rebuilt the foundation the whole thing sits on — most of which you couldn’t see, and that’s the part I’m most careful about.
Your AI, your provider — per tier
For a long time “which model” really meant “which one of a single vendor’s brand names.” In June that opened up. Model tiers are now named by what they do — fast, balanced, deep — and each tier can run on the provider and model you choose. Anthropic and Mistral are both first-class now (the old “EU-sovereign mode” toggle is gone — it’s just a provider choice), and in the opt-in Hybrid mode you can mix them: a fast Mistral tier for cheap background work alongside an Anthropic deep tier for the hard thinking, in the same session. Your included budget caps your spend, not an artificial tier ceiling. Your tools, your keys, your provider — now down to the individual model.
The whole product became the chat
This is the change I care about most. Instead of building a bespoke form for every capability, June moved them into the one surface that already understands context: the conversation.
Saved workflows became a real product surface (v1.17): schedule one on a cron, get pulled into an unread chat thread when a run needs your input, and edit or fix a workflow by talking to the agent — no step-editor. The inbox went chat-first (v1.18): replying to or composing a mail opens a chat with the message loaded as context, and the old standalone composer is gone. And the navigation itself was reorganised to map to the agent loop (v1.19) — Workflows, Triggers, Tasks, with “open in chat” actions on tasks, contacts, and API profiles. Every old bookmark still redirects, so nothing broke.
The bet underneath all of it: a chat that carries context beats a hand-built screen for almost everything.
Documents and mailboxes it can actually reach
Two practical reach extensions. A PDF or Word file dropped into chat is now read — text-extracted server-side — instead of rejected, and what it read persists into memory so the document comes back in later conversations (v1.20). And you can connect a mailbox straight from chat (v1.21): the agent resolves the provider’s servers and hands off to a consent step where you type the app password — and that password goes from your browser to the vault directly. It never enters the agent’s context, the run history, or the prompt store. The agent only learns whether the connection worked.
The foundation you couldn’t see
Here’s the honest part: the single biggest piece of work in June is the one you can’t point at in the UI. Under the name Foundation Rework (v1.22) I shipped a new consolidated per-tenant data store — one place for people, organisations, memories, relationships, and the workflow/trigger/task layer, instead of the several overlapping databases that had grown over time. It shipped dormant, behind a flag that’s off by default, doing nothing to your instance yet.
That sounds anticlimactic, on purpose. A foundation you rebuild under a running product has to be provably safe before it carries any weight. Shipping it inert, one careful slice at a time, is what made the 2.0 release in early July possible — the one where it actually goes live. Most of the month went here, quietly.
The unglamorous layer underneath
The rest was the unglamorous layer you should expect from software that hosts your data. Clearer provenance — the agent can tell you where a given fact came from (v1.12). Cost-transparency tooling and a thread debug-export, so you can see exactly what a long conversation cost and why (v1.16). And a floor of compliance and safety work — sanctions handling, an AI-transparency disclosure, and a running wave of security hardening.
That was June. It was preparation, mostly — and in the first days of July the foundation went live as 2.0. If you read this far, you’re still exactly who I write these for. The granular history is in the changelog, and next month there’ll be another one.