Invitations expire after 7 days and include a single-use link. The console shows pending invitations and lets you revoke them.

Use the console or `DELETE /v1/users/me/workspaces/:id`. You cannot leave if you are the last owner.

The API Keys page has a Connected MCP applications card with a Disconnect button for each client.

Use the console or `DELETE /v1/skills/:id`. A published skill leaves the published skills feed, but agents that already downloaded it keep their copy. The delete is refused if another skill calls it or a distillation run is in flight.

The credential test and every extraction check that the endpoint is public, that redirects stay public, and that the region is valid. Calls to internal addresses are refused.

The test and switch endpoints now refuse NAMS hosts and the ops host.

Each user gets 10 attempts per 15 minutes across all targets. This prevents a workspace admin from locking NAMS out of a database by aiming wrong-password logins at it.

A user joins a workspace only when they accept an invitation from an owner or admin. Members added before this release who have not yet signed up will lose access until they accept the invitation in the console.

You see a consent page once per client, showing the client name, redirect host, scopes, and your workspaces. You can deny the request. Approving a client again ends tokens from any earlier approval.

An MCP token or a narrowed API key cannot access a key with wider scopes. Your Default key now works only in your first workspace.

Model load time dropped from 16.7 seconds to 8.6–10.8 seconds.

Before, only workspace name conflicts showed their message; other conflicts fell back to a generic error.

The consent page can now receive Allow and Deny requests from the browser.

Long messages now receive embeddings and are visible to search.

The console shows the import date for imported snapshots until they expire.

If a copy operation fails because the database temporarily stopped accepting writes, it will automatically retry for up to a minute instead of failing immediately.

When a database becomes unavailable, the migration process now recognizes this and stops trying to migrate it, eliminating the error logs that were appearing every few minutes.

Workspaces created before the managed-database pool had their data backed up by the legacy service. Those backups are now imported as restorable snapshots, so you can restore them from the console the same way as any other snapshot.

When the database becomes temporarily unreachable, requests are retried instead of failing immediately. You'll see fewer transient errors in your logs.

This reduces unnecessary work and makes review processing more efficient.

You can now remove all of a specific user's messages, relationships, and metadata from a workspace in one operation.

This improves isolation and performance when managing conversations across multiple users in the same workspace.

This provides more reliable state management in the web console.

You can now distinguish between service issues and workspace-specific problems.

You can attach and retrieve custom metadata alongside each message.

List endpoints now support filtering by userId and cursor-based pagination for better control over large result sets.

Entity relationships are now fully queryable with consistent identifiers and directional information.

Deletion now supports an option to remove all associated extraction data.

Unknown fields and data after the JSON body are now rejected on conversation, message, search, and query endpoints.

Embedding now respects both character and token limits, so no message is silently dropped from vector search.

The service enforces your deadline across all enrichment attempts, so abandoned calls no longer consume resources after you stop waiting.

Entity and relationship extraction now run more efficiently.

Duplicate detection now runs on all extracted entities, and the graph no longer loses relationships in large batches.

Unprovisioned tenants, DNS failures on reaped databases, and caller-supplied query errors are now logged at the appropriate level.

Your workspace ID is carried through all extraction calls so activity is traceable.

Deletion now removes messages, entities, relationships, and all associated metadata.

The message list endpoint now returns results in reverse chronological order.

Blank names are now rejected instead of being silently embedded.

When a write matches multiple edges, the lowest ID is now returned consistently.

API documentation is now generated from annotations, so it stays in sync with the code.

Results over 1000 rows are now rejected instead of being truncated silently.

The provisioning flow now adopts an existing database instead of spinning forever.

Memory usage stays under control as the deferred stream no longer fills to capacity.