Skip to content

Admin Dashboard: Resilience

Resilience shows one circuit breaker per MCP server. Breakers stop repeated calls to a failing dependency from degrading every agent turn.

  • Closed: calls flow normally.
  • Open: consecutive failures crossed the threshold, so calls fail fast.
  • Half-open: the cooldown elapsed and a probe is testing recovery.

The page polls every five seconds and shows failure count and the most recent failure time.

  1. Fix or verify the upstream MCP server first.
  2. Open Resilience and locate its breaker.
  3. Choose Reset only for an open or half-open breaker.
  4. Confirm the state returns to Closed, then run a representative agent tool call.

The expected result is a closed breaker and a successful tool call. Resetting does not repair authentication, transport, catalog, or upstream failures.

Use MCP Servers to refresh tools/list and inspect connection status. Use Tool Call Log for arguments, timing, and result status. When a live SSE transport disconnects, pending requests fail immediately with a safe error and the transport reconnects. If a server sends an SSE event larger than 1 MiB, reduce or split its response so each event stays within the supported size.