Tools that fit your hand — a cleaner way to connect and use your apps

This release is about the seam where your assistant meets your apps — connecting them, choosing what it can touch, and trusting what it does. It got cramped. We opened it up.

A tool picker you can actually see

Adding apps to an agent used to happen in a tiny dropdown that was hard to search and easy to lose. It's now a full-height panel that slides in from the side, with room to breathe and three clear tabs:

• Your apps — pin a whole connection so the agent can use everything in it.
• Specific actions — pin single actions instead, when you want a tighter scope. Filter them by read-only or makes changes, and hit Add all read-only to bundle every safe action in one tap.
• Add new app — search the whole library and connect something new right there, without leaving the agent you're building.

gmail · search inboxread-only — ran automatically
gmail · send emailmakes changes — held for approval
Draft ready — send it to the team?Not yetSend

That read-only / makes-changes split isn't just a label. In Safe mode, actions that only read run on their own, while anything that changes something waits for your tap — so "add all read-only" gives a capable agent that genuinely can't touch anything on its own.

Connections moved into Settings

Connecting apps is something you set up once and rarely revisit, so it no longer needs a permanent spot in the sidebar. It now lives in Settings → Connections — the same full library and connection manager, just where account-level things belong. The sidebar is shorter and calmer for it. And because you can now connect a new app from inside a chat or an agent, you rarely need to go there at all.

Search models now show their sources

Some models do their own web search and hand back an answer with citations baked in. Those citations weren't being surfaced — the sources were there, we just weren't showing them. Now they land in the same tidy citations block as every other web result, so you can check a search-model answer as easily as any other.

Also shipped and fixed

• Handle auto-fills from an agent's name as you type — stops the moment you edit it yourself.
• The plan/review trace in adversarial runs no longer mislabels a "revisions requested" verdict as a failure — the reviewer doing its job reads as success, with the critique in the trace.
• Faster, cheaper thinking behind the scenes. The internal model that plans tool use, routes agents, and drafts skills now routes over a small pool of low-cost models, cheapest-first — invisible to you, lighter on your allowance.
• Mobile and installed-app polish — the picker, controls, and every settings screen now behave on a phone, and the home-screen app respects the notch and safe areas.
• The app icon is back to the 1Chat mark on the tab and home screen.

As always, if something feels off, Report an issue from the sidebar (or the menu on mobile) goes straight to us — a good chunk of this list started there.