n8n

Across the past month, the release cadence has focused on targeted fixes spanning three active maintenance trains (2.35.x, 2.36.x, and 1.123.x).

The AI Assistant handles larger models without failing verification. The model verification token limit was raised across both the 2.35 and 2.36 trains, resolving failures that occurred when checking AI credentials for larger context models. Agent preview conversations also keep user messages correctly right-aligned in the editor.

End-user credential resolution is more reliable. Chat and MCP trigger auth modes are now validated against end-user credentials rather than skipping that check, and node parameters now correctly resolve end-user credentials when the node loads — both fixes apply to builds running multi-user workflows.

Expression engine initialization is scoped more tightly. The expression engine no longer initializes for CLI commands that don't evaluate expressions, reducing startup overhead for non-evaluation commands. Test-webhook isolates are released only after full teardown completes, and a trigger's closeFunction is wrapped in an expression isolate at creation time so teardown is always scoped correctly.

Slow task runners no longer trigger unnecessary restarts. Previously, a runner that fell behind on processing could be killed and restarted even when it was still making progress. The fix distinguishes between a slow runner and an unresponsive one, keeping long-running executions alive.

Environment variable parsing is consistent across trains. Env values are normalized before schema-based parsing, preventing type coercion failures that could misread boolean or numeric env vars.

The Instance AI landing page now surfaces the personalization survey and community registration modal on first visit.