We are implementing an integration with the official Atlassian Rovo MCP Server.
OAuth discovery and dynamic client registration succeed, and the consent screen opens correctly. After selecting our Jira test site and clicking Accept, Atlassian returns:
invalid_request — Incorrect request parameters
The browser never reaches the registered localhost callback, so no authorization code or tokens are obtained.
We independently reproduced this on 9 October 2026 using Python MCP SDK 2.3.0, with no application code or database involved:
No Write scope, offline_access, or prompt parameter
Consent showed the expected client and Read/Search permissions
No callback received or token exchange attempted
The test user has Jira access, and the inspected MCP permission and data-security settings allow access.
Could an Atlassian team member help identify whether this is a client-metadata compatibility issue or an account/site condition rejected during consent? We have an earlier Atlassian request ID and timestamp available for investigation through an appropriate private channel.
Scopes and resource match what https://mcp.atlassian.com/.well-known/oauth-protected-resource/v2/mcp publishes today. A domain block shows its own admin message on the consent screen, and for a localhost callback the fix is an admin allowlisting the exact redirect URL under Domain settings. Probably not that.
Two authorization servers, though. The root /.well-known/oauth-authorization-server still lists /v1/authorize and /v1/register. The v2 resource metadata points at https://auth.atlassian.com/VCeDsk8ZHncYF1g234fKtc4lNipbBhu3, whose own metadata registers at .../dcr/register. Which one did the SDK register with? The URL that returned the 201 shows it. April’s changelog entry on the new DCR provider says a client_id or discovery document from the old one won’t be recognised.
Couldn’t find anything public on invalid_request at Accept. Registered on the new server? Then the request ID goes in a developer support ticket.
We encountered the same error after clicking Accept, before reaching our callback.
Chrome DevTools revealed that Atlassian’s consent form submitted read:me, read:account, and offline_access in addition to our requested tool scopes.
We resolved it by registering a new OAuth client with those three additional scopes, including them in the authorization request, and registering both authorization_code and refresh_token grant types. We retained our selected agent-interface tool scopes.
Login then succeeded. We changed these settings together, so we haven’t isolated which individual scope or grant was essential, but this combination fixed our case.