Skip to main contentPerformance Optimization
X21 is tuned for responsiveness, but complex workbooks can still slow things down. Follow these tips to keep interactions snappy.
- Batch writes – Use a single
write_values call for multi-row updates instead of multiple small calls. Each call incurs Excel interop overhead.
- Limit formatting changes – Apply formatting once on the final range rather than per column to reduce COM traffic.
- Reuse context – When running consecutive operations, keep the same workbook selection so Claude doesn’t need to re-read large ranges.
Manage Conversation Size
- Encourage concise prompts and responses. Excessive narration increases token usage, slowing Claude’s replies.
- When you’re done with a topic, restart the conversation. This trims the history and avoids triggering compacting unnecessarily.
- If Claude starts summarising repeatedly, ask it to reset the plan or provide only the final results.
Attachment Guidelines
- Compress PDFs or limit them to relevant pages before attaching. Oversized documents increase WebSocket payloads and model processing time.
- Use images only when necessary; textual data is cheaper to process and faster to reason about.
Excel-Specific Tips
- Close other heavy add-ins to free up Excel’s single-threaded automation pipeline.
- Avoid running X21 while long-running formulas, recalculations, or external data refreshes are in flight.
- Keep workbook size reasonable. If datasets exceed a few hundred thousand cells, offload intermediate steps to Power Query or a database.
Backend Monitoring
- Check
%LOCALAPPDATA%\X21\X21-deno\Logs for slow tool execution metrics. Langfuse spans also display durations per tool.
- If the backend becomes unresponsive, restart the add-in (close Excel). The Deno server shuts down automatically when the process exits.
- Monitor network security tools that might inspect localhost traffic; excessive scanning can degrade WebSocket performance.
By batching operations, trimming conversations, and keeping workbooks healthy, you’ll see streaming responses land in seconds rather than minutes.