Slow Calls

Slow tool calls are the most common source of poor user experience in MCP servers. A tool that takes 2 seconds instead of 50ms makes the AI client feel unresponsive — and the user has no visibility into why.

The Slow Calls tab on the server dashboard surfaces exactly which calls were slow, how slow they were, and gives you the data to figure out why.

Reading the table

Each row is a single MCP tool call, ranked by total latency (slowest first):

Column	What it shows
#	Rank by latency
Time	When the call happened
Tool	Which tool was called (clickable — jumps to Event Log filtered to that tool)
Latency	Total end-to-end time (proxy receive → response sent). Red when notably slow
Upstream	Time your MCP server spent processing. Compare with Latency to see proxy overhead
Client	Which AI client made the call (e.g. mcpr-studio, claude-ai, openai-mcp)
Session	Session ID (clickable — jumps to Sessions tab to see the full conversation)
Error	Error message, if the call also failed
Size	Response body size — large payloads can indicate the cause

What to look for

Upstream vs Latency gap

If Upstream is close to Latency, your MCP server is the bottleneck — the proxy added negligible overhead. If there’s a gap, the proxy or network added time (rare, but check for large response bodies).

Large response sizes

A call returning 9.5 KB is doing more work than one returning 550 B. If a slow call has a large size, the tool may be fetching too much data or returning verbose responses. Consider pagination or trimming the response.

Errors on slow calls

A slow call that also has an error often points to a timeout. Your MCP server spent time working, then failed. Check if the upstream service has its own timeout that’s shorter than expected.

Repeated tools

If the same tool appears multiple times in the slow calls list (like review_vocab in the screenshot), it may have inconsistent performance — sometimes fast, sometimes slow. This often indicates:

• Cache misses — fast when cached, slow when not
• Variable query complexity — depends on input parameters
• External dependency jitter — database or API response time varies

Client patterns

Different clients may trigger different performance characteristics. If openai-mcp calls are consistently slower than mcpr-studio calls for the same tool, investigate whether the client sends different parameters or triggers different code paths.

Page size

Use the Show toggle (10 / 20 / 50 / 100) in the top right to control how many slow calls are displayed. Start with 20 to get the big picture, then increase to 100 when investigating a specific tool.

Debugging workflow

1. Open Slow Calls — identify which tools are slow and how slow
2. Check Upstream vs Latency — is it your server or the proxy?
3. Check Size — is the response unusually large?
4. Click the Session ID — see what happened before and after. Was the client retrying? Were other tools also slow in the same session?
5. Click the Tool name — jump to Event Log filtered to that tool. See if the slowness is consistent or an outlier
6. Cross-reference with Latency tab — check if the tool’s p95 is high in general, or if this was a one-off spike

Relation to other tabs

Tab	How it complements Slow Calls
Latency	Shows percentile trends over time — “is this getting worse?”
Tool Health	Shows per-tool p50/p95/p99 — “what’s normal for this tool?”
Event Log	Full request details — “what exactly was sent and returned?”
Sessions	Conversation context — “what happened around this slow call?”
Errors	Correlate — “are slow calls also failing?”