// November 10, 2026v1.51.3
Platform
Identity
Editing an agent used to reopen the full six-step setup wizard just to change which servers it reaches. Agents now get their own MCP sign-in tab, second after Overview, where adding, removing, or reconnecting a server saves immediately with a toast, no wizard and no Save button. The old Credentials tab is renamed Agent sign-in and issues a new key through that same flow instead of asking you to pick servers again. The Permissions tab now leads with the MCP servers an agent can reach, in their own section above its other grants, and its "Acting on behalf of" panel lists every server an agent may connect to alongside whose credential it uses there. Platform MCP can now read all of this too, through
get_agent_mcp_access and list_agent_authorization_checks. More on this in Authorizing agents and Platform MCP.
Features
- Change an agent's MCP servers and sign-in from its own tab #8812 - A new MCP sign-in tab lists an agent's servers and how it signs in to each, and saves every change as it's made instead of walking the setup wizard again. A red count shows how many servers still need someone to sign in, and the old "Review connections" and "Change servers" links now land here. (Author: @mfbx9da4)
- Issue an agent key from a renamed Agent sign-in tab #8809 - An agent's Credentials tab is now called Agent sign-in and shows the same sign-in diagram as the creation wizard. Issuing a key there runs the wizard's own sign-in, install, and verify steps for the servers the agent already reaches, instead of asking you to pick servers or set an expiry again. (Author: @mfbx9da4)
- An agent's Permissions tab now leads with its MCP servers #8706 - The servers an agent may connect to now get their own section at the top of its Permissions tab, with an Edit servers button that opens the agent's MCP server step directly. (Author: @mfbx9da4)
- See exactly which credential an agent uses on each server, and how often #8713 - An agent's "Acting on behalf of" panel now lists every MCP server it may connect to, which credential it uses there, and its calls in the period, so a server it can reach but never called shows plainly as zero calls. (Author: @mfbx9da4)
- Platform MCP can now read an agent's access and authorization checks #8758 -
get_agent_mcp_accesslists the MCP servers an agent identity may connect to, the rules behind each, and whether it signs in with a shared credential or on behalf of named people.list_agent_authorization_checkslists its recent checks, denials first. (Author: @mfbx9da4)
Bug fixes
- A broken agent connection now says exactly what's wrong and who fixes it #8712 - Instead of a generic "ask your administrator," a broken connection now says the server needs re-authentication and links whoever can edit the agent straight to its MCP sign-in step. A rejected upstream call, a permissions problem, and a shared credential issue each get their own message naming no person and carrying no credential. (Author: @mfbx9da4)
- A gateway tool call blocked by a sign-in change now says who can fix it #8733 - When an admin changes a gateway server's sign-in setup after you approved it, the error now names the server and says whether reconnecting helps or an administrator has to step in, instead of always saying "reconnect." You can also now pick between two connected services for one server when sign-in enforcement is off. (Author: @daviddanialy)
- An agent's allowed MCP connections now show up next to its denials #8719 - Authorization checks now list failed checks first, grouped by scope and resource, followed by a section of the servers the agent was allowed to connect to, so a real denial no longer gets buried among routine allowed checks. (Author: @mfbx9da4)
- Editing an agent's permissions no longer silently drops grants it can't display #8707 - Edit permissions now loads every grant exactly as stored, including grants on a single MCP server or a whole project, and saves only the scopes you actually changed. The dialog lists every grant a save will add or remove before you confirm. (Author: @mfbx9da4)
- Every MCP server grant on an agent's Permissions tab now names its project #8724 - Grants saved from the permission picker without a project now read it from the server, so the Permissions tab never shows a grant with no project attached. (Author: @mfbx9da4)
- Opening a session from an agent's page no longer loses your place #8716 - Opening a session from an agent's Overview or Sessions tab now opens it in a drawer over the identity page, keeping the sidebar and session list in place, and the open session is kept in the URL so it can be linked, or closed with back. (Author: @mfbx9da4)
- An agent's tool calls chart now splits out failed and blocked calls #8726 - The chart now colors calls by outcome, succeeded, failed, blocked, or pending, with a legend and a failed count, using the same colors as the Tool logs status dots. (Author: @mfbx9da4)
- Grants on a hosted MCP server now link to the server itself #8725 - An agent's Permissions tab and its authorization checks now link a hosted server's grants to its
/mcppage instead of an old/toolsets/redirect. (Author: @mfbx9da4) - Tool logs now break out a separate header for each day #8714 - Project Tool logs and an identity's Tool logs now open every day in local time with its own date header, so earlier rows no longer sit under the first day's date. (Author: @mfbx9da4)
- Long agent sessions now show their real duration #8715 - A session of a day or more now shows its duration in days and hours, such as "3d 20h," in session lists, the Costs session table, the session CSV, and turn cost badges. (Author: @mfbx9da4)
- An identity's breadcrumb now matches its tab names #8721 - The breadcrumb now names a section the way its tab does, such as "Controls & Safety" on an agent, instead of title-casing the URL. (Author: @mfbx9da4)
- Every resource picker checkbox now has its own accessible name #8723 - Screen readers can now tell rows in the resource picker apart, including a pinned, already-selected row. (Author: @mfbx9da4)
- Tool usage on a busy project's identity page now loads reliably #8709 - Tool usage reads now find an agent's traces before grouping them, so they finish within the time limit on busy projects instead of returning zero calls, and Tool logs now say when a read failed with a Retry instead of looping silently. (Author: @mfbx9da4)
- An agent's authorization checks no longer get buried under hook ingestion noise #8717 - Hook ingestion checks, one per hook event, are now left out of an agent's authorization checks, so the checks that actually say what the agent may reach aren't drowned out. (Author: @mfbx9da4)
