Release notes
What has shipped, newest first.
The Elevate 1.0 improvements land here, starting with cost numbers you can trust.
Fixed- Spending reports could show a wildly inflated figure for “the last hour” — one workspace was told it had spent $55.87 in an hour when the real figure was $10.01. A long-running conversation had its entire history re-dated onto the current hour every time the numbers refreshed. Spend is now recorded once, when it happens, so an hour means that hour and history stops moving underneath you.
- An owner of a workspace could not add a workspace-wide connector — it refused with a permissions error even though they had permission. They can now.
- Saving a credential could quietly produce one that could never be read back. It reported success and then failed later, somewhere unrelated. It now refuses at the point of saving instead.
- Scheduled work used to show up as spending nobody incurred — on one day that was a third of the bill. It is now attributed to the project and the person it belongs to.
New- Cost can be broken down per project and per person, with the real project names, so a report cannot quietly attach one project’s spending to another’s name.
- You can run this workspace on models other than Claude — OpenAI and Google are now selectable, with your own key.
- You can put your own Claude subscription behind your direct messages, instead of the workspace’s metered billing.
Changed- Conversations are cheaper to run. Oversized tool results are trimmed, long conversations start summarising themselves much earlier, an agent that gets stuck repeating itself is now stopped, and tools no workspace uses are no longer offered. Nothing you asked for is removed — the same answers cost less to reach.
- Very large skills are loaded a section at a time instead of whole. The content is identical; it just stops riding along in every later step of the conversation.
Updates to an existing workspace apply cleanly.
Fixed- A workspace that had been running for a while could, right after an update, start refusing actions — creating a project or adding a connector would fail, and the record of who did what stopped filling in. Updating now brings the workspace’s own records up to date automatically as it starts, so the first thing you do after an update just works, the same as on a brand-new workspace.
Hard to miss where Elevate answered you.
Changed- Clicking “+ New project”, “+ Create agent” or “+ Add connector” used to answer in a small card in the bottom corner, while you were looking at the top of the page — easy to miss entirely. The page behind now dims and the panel grows, so it is obvious where the conversation went, and the cursor is already in the box waiting for your answer.
- The dim is a signpost, not a trap. Reply, press Escape, or click anywhere on the dimmed page and it clears — the conversation stays open underneath, and you can carry on looking around while Elevate works.
- Clicking the pill yourself does not dim anything. You already know where the panel is.
Creating things is a conversation with Elevate now, not a form.
Changed- “+ New project”, “+ Create agent” and “+ Add connector” no longer open a dialog box. They open Elevate on the right with the request already made, and it asks you what it needs — the same way the cards on the front door work.
- The old New project dialog took a name and made an empty folder. A project is only worth having once the right accounts are bound to it and the right agents are working in it, and Elevate now sorts that out with you in one go rather than leaving you three more screens to find.
- Adding a connector from the Connectors screen used to make it personal to you whatever you intended — that was the only option the form had. Elevate asks where it should live: yours alone, one project’s, or the whole workspace’s. Entering the credential itself is unchanged; the form appears in the chat and the secret still goes straight to the vault.
- “+ Create agent” used to fail for members with a permission error after they had filled the whole form in. Ask Elevate for an agent and it does the right thing for you either way — an admin gets the agent, a member gets a request an admin can approve.
Your personal connectors are yours again.
Fixed- The Connectors screen was showing everyone in the workspace the personal connectors belonging to whoever had set one up — listed under “Personal · Yours alone”, as though they were your own. Nobody could ever use someone else’s (the credential itself stayed sealed), but the names of what a colleague had connected privately were on the page. You now see only your own.
- The same list sent to Slack or Teams by `@knet connector list` was showing every member the same thing. It is now filtered the same way.
- Connectors shared with the whole workspace, and a project’s own connectors, are unchanged — they are meant to be seen by the people who share them.
A What’s New reel, so you can catch up in a minute.
New- The release notes page now opens with the last couple of big things Elevate learned to do, written as what you can now do rather than what changed in which build. Scroll or swipe between them; the full build-by-build list is still underneath.
- Each one says whether your workspace already has it, because every workspace updates on its own schedule.
Just ask Elevate what Claude can reach.
New- You can now ask Elevate directly — “what can Claude reach here?”, or “only let Claude see the Engage tools” — instead of needing the chat command. Anyone can ask what is exposed; only admins and owners can change it, and asking Elevate does not get you around that.
- This also means the setting no longer depends on having Slack or Teams connected. Workspaces using only the web app had no way to reach it at all.
Anyone can check what Claude reaches; only admins can change it.
Fixed- Asking `@knet admin mcp tools` as a member answered “Only admins can change what the MCP connector exposes” — a refusal for something you had not asked to do. Anyone in the org can now see the list. Changing it is still admins only.
You decide which of your tools Claude can reach.
New- Connecting Elevate to Claude used to offer every tool your workspace had installed, all or nothing. An admin can now pick the list. Send `@knet admin mcp tools` to see what is available and what is currently on, `@knet admin mcp tools set` followed by the ones you want, and `@knet admin mcp tools all` to turn everything back on.
- Changes take effect on the next thing you ask Claude. Nobody has to reconnect, nothing restarts, and people already using the connector simply stop seeing the tools you switched off.
- Narrowing the list is no longer a choice between exposing a tool and uninstalling it — which would also have taken it away from Slack and Teams, where the same tools are still in use.
- `@knet admin mcp tools set none` switches the connector off entirely while leaving it connected, which is the fastest thing to reach for if you are unsure and want to stop first and think second.
- Mistype a tool name and the whole change is refused, with the real names listed. It does not apply the part it understood — that would leave you exposing one tool fewer than you meant, with nothing to indicate it.
Changed- The list is for the whole organisation, not per person. What each person can actually do with a tool still follows their own role, checked fresh on every request — so someone whose access you reduce loses it on their next question, not whenever their sign-in happens to expire.
- Only admins and owners can change the list. Members are refused, and nothing is changed by the attempt.
- Every change is recorded in the audit log with who made it, because it silently changes what everyone already connected can reach.
Choose what Claude can reach, and which project it works on.
New- An admin can now list exactly which tools the Claude connector offers, instead of it offering everything installed. Set it once and it applies immediately — no restart, and nothing needs uninstalling.
- Tools that belong to a project can now be pointed at one from a Claude conversation. Name the project and Claude looks up the rest; before this, every such call failed.
Signing in through Google, or from Claude, now sticks.
Fixed- Signing in with Google appeared to work and then returned you to the sign-in page. The session was real; your browser was holding it back because of how the sign-in cookie was scoped. It is now sent when you arrive from Google or from Claude.
Signing in no longer abandons a connector setup.
Fixed- Connecting Elevate to Claude sent you to the sign-in page and then dropped you on your dashboard, so the connection was never finished. Sign-in now returns you to where you were and completes it.
Use your Elevate skills from Claude.
New- Elevate can now be added to Claude as a connector, so the skills your workspace already has show up as tools you can use directly from a Claude conversation. You sign in with the same Elevate account you use here, and Claude acts as you — the same permissions, the same audit trail.
- Only the skills your workspace has installed are offered. Administration — users, secrets, fleet settings — is never reachable this way.
- A long-running job started from Claude now reports back when it finishes, so you can see the result where you started it instead of only in the chat channel it was posted to.
Your reports now reach the people who can fix them.
Fixed- Reports were being captured but had no way to travel to the engineering team, so they sat where they were filed. They are now collected and raised as tracked work, and a report keeps a link to the ticket it became.
- The same problem reported on more than one workspace is recognised as one problem affecting several, rather than being counted separately.
Tell Elevate when something is broken.
New- Elevate now takes bug reports and feature requests about the platform itself. Say what went wrong and it files it — you do not need to find a form.
- Reports about the same problem are collapsed into one entry with a count of how many people hit it. That count is how things get prioritised, so reporting something already reported is still worth doing.
- A bug that stops happening after an upgrade closes itself, and reopens if it comes back.
- Admins decide which reports become work. Everyone can report; not every report changes the product.
Changed- Reports capture the version, the failing tool and its error — never your message text, project names, connector names or customer names.
Elevate knows which conversation it is in.
Fixed- Elevate could tell you it had moved you into a project when it had not. It now says plainly that a chat with Elevate stays outside projects, and points you to the project chat instead.
- An agent Elevate created could be seated in a different conversation than the one that asked for it. Elevate was working out which chat it was in by guessing; now it is told.
- Elevate could report a step as done when the step had actually been refused.
- A chat with Elevate in it could, in one case, be filed into a project — which is not a thing that is supposed to be possible.
Changed- Asked to staff ongoing work, Elevate now looks at who already works here and offers to assign them, rather than inventing a new agent every time. If it does propose someone new, it has to say which existing agent it considered.
Elevate builds the team. The team does the work.
New- New Chat is a conversation with Elevate about setting work up. Ask for something to be done and it creates or assigns the agent that will do it, then hands over.
- Projects are where collaboration happens: a project chat holds everyone on the project plus the agents assigned to it. Address one with @.
- Assigning an agent to a project is what gives it that project's connectors — so the same agent works on many projects and picks up each one's access in turn.
Changed- An agent is personal or project, never both. A project agent can never reach a personal connector — not by assignment, not by ownership, not by an administrator.
- Elevate is never in a project chat, and a chat with Elevate holds one person.
- When you speak to an agent, it acts as YOU. Its owner is who it reports to, not an identity it borrows — and the cost lands on whoever spoke.