Your Assistant, Available in Any AI Client
Your workspace assistant now holds real conversations from inside Claude, ChatGPT, Cursor or your own code — with a connection page rebuilt to match.
Until now, connecting an outside AI client to your workspace meant settling for a handful of narrow lookup tools — fine for fetching a fact, useless for actual thinking. That changes today. You can now talk to your assistant from anywhere, using a new conversational capability on your workspace connection. Ask a question from claude.ai, ChatGPT, Claude Desktop, Cursor, or a script you wrote yourself, and you get the same assistant you'd get in the app: your memory, your company knowledge, your connected integrations, and web search, all working together on a real answer.
Conversations behave like conversations. Every reply comes back with a conversation identifier, and passing it along continues the same exchange, so a follow-up question doesn't start from scratch. Each conversation stays scoped to your workspace, the client that started it, and the member whose credentials are in play — nobody sees anyone else's thread. Messages are priced like ordinary assistant chat, one credit each, with a rolling daily ceiling per workspace so an automated caller left running overnight can't quietly spend your month. And a read-only credential now genuinely means read-only: memory and automation changes are withheld outright, not merely discouraged.
The connection page in Settings → API access has been rebuilt around the only question that actually matters — what kind of client are you connecting? Three self-contained paths now cover connecting in a browser with a single sign-in, connecting a desktop app with a token and config file, and connecting your own code. That last path is new: a copyable request to start from, an explanation of how conversation continuity works, a pre-filled command to verify your connection, and a clear warning to allow at least five minutes per request, since a thoughtful answer can take a while. Access tokens now appear only inside the two paths that need one, so there's no longer a step to figure out and skip.
The same page is now honest about what it offers. It lists every capability your connection actually serves, generated from the live source rather than a handwritten list that had drifted out of date — a new tool shows up automatically, described in plain language. Credit costs are stated up front, and connected apps show whether they can read only or read and make changes, instead of a raw permission string. Alongside all this, we fixed a defect that could break long-running conversations after roughly ten exchanges while still charging for them — a fix that now applies consistently across every surface your assistant runs on — tidied up how connection failures are reported so errors appear beside the control that caused them, and corrected a faulty connection probe that could send some clients into an endless retry loop.
If you've been waiting to bring your workspace assistant into the tools you already work in, this is the release to try it. Open Settings → API access, pick the path that matches your client, and start a conversation.