ClickHouse

@ClickHouse show me the schema for the events table in the analytics database, including a few sample rows so I can understand the data structure

The `analytics.events` table has 8 columns: `event_id` (UUID), `user_id` (UInt64), `event_type` (String), `timestamp` (DateTime), `properties` (String as JSON), `session_id` (String), `device_type` (Enum8), and `country_code` (FixedString(2)). The table uses the MergeTree engine partitioned by toYYYYMM(timestamp) and ordered by (user_id, timestamp). Sample data shows events like 'page_view', 'button_click', and 'purchase' with timestamps from the last 24 hours. Total row count: 847M rows, compressed size: 12.3 GB.

This example demonstrates schema introspection with sample data, which helps the AI construct accurate queries without guessing column names or types. Requires read access to the target database. Sample data is limited to a few rows to avoid overwhelming the context window with large result sets.

@ClickHouse run a query to find the top 10 countries by user signups in the last 7 days, grouped by day, from the users table

Query executed successfully in 0.34 seconds. Results show the United States leading with 3,847 signups (peak on March 15 with 612), followed by United Kingdom (1,923), Germany (1,456), France (1,203), and Canada (1,087). Daily trends indicate a 23% week-over-week increase in US signups, while European countries show steadier growth around 8-12%. The query scanned 2.1M rows and returned 70 result rows (10 countries × 7 days).

This showcases ClickHouse's strength in analytical queries over large datasets with fast aggregation. The AI translated natural language into optimized SQL with proper date filtering and grouping. Be aware that complex queries on very large tables may hit execution timeouts—consider adding LIMIT clauses or narrower time windows for exploratory queries.

@ClickHouse analyze our page load performance: calculate the median, 95th percentile, and 99th percentile load times by device type for the past 30 days, and tell me which segments need attention

Performance analysis complete. Desktop users see median load time of 1.2s (p95: 3.1s, p99: 5.8s). Mobile users show median 2.4s (p95: 6.7s, p99: 11.2s). Tablet users median 1.8s (p95: 4.9s, p99: 8.3s). The data reveals mobile users experience 2× slower median loads and significantly worse tail latency. The p99 for mobile (11.2s) suggests ~1% of mobile sessions face severe performance issues—likely correlated with older devices or poor network conditions based on the distribution shape. Recommend investigating mobile-specific optimizations and considering separate performance budgets per device class.

This example pairs ClickHouse's quantile functions with AI reasoning to surface actionable insights from raw performance data. The AI constructs statistically appropriate queries (using quantileExact or quantileTiming functions) and interprets the results in business context. Requires the MCP to have access to tables with timestamp and performance metric columns.

When ClickHouse MCP beats manual SQL for weekly metrics

A 6-person product team runs weekly sprint reviews where they pull user funnel metrics, feature adoption rates, and performance benchmarks from ClickHouse. Without this MCP, someone writes SQL in the ClickHouse console, exports CSVs, and pastes numbers into slides. With the MCP, the team asks Switchy to pull last week's conversion rates or compare feature usage across cohorts—Switchy executes the query, formats the results, and drops them into the shared workspace in under 10 seconds. The trade-off: if your queries need custom UDFs or complex window functions that require iterative tuning, you'll still want a SQL IDE open. But for standard aggregations and time-series pulls during live meetings, this MCP eliminates the context-switch tax. Worth it if your team references ClickHouse data more than twice a week.

Using ClickHouse MCP to surface error patterns in real time

A 3-person support team at a SaaS company uses ClickHouse to log API errors and user events. When a customer reports a bug, the team needs to check if it's isolated or part of a broader pattern—usually by asking engineering to run queries. With this MCP in Switchy, support can ask 'show me all 500 errors for customer X in the last 48 hours' or 'count failed logins by endpoint today' without leaving the ticket thread. The MCP lists databases, pulls table schemas, and executes the query in one pass. The limit: if your ClickHouse instance has 200+ tables or requires VPN access, OAuth2 setup gets fiddly and schema discovery slows down. Best fit for teams with under 50 tables who need ad-hoc log queries multiple times per day.

When ClickHouse MCP replaces dashboard fatigue for small teams

A 4-person growth team runs paid campaigns across three channels and stores event data in ClickHouse. They have dashboards, but answering one-off questions—'did iOS conversions drop after yesterday's push?'—means waiting for a data analyst or building a new chart. With this MCP, the team asks Switchy to query campaign performance by device, compare week-over-week spend efficiency, or pull attribution breakdowns on demand. The MCP handles schema lookups and query execution in seconds, so the team gets answers in Slack instead of waiting for dashboard updates. The boundary: if your campaigns generate more than 10 million events per day, query latency spikes and you'll want pre-aggregated tables. This MCP shines for teams under 5 people who need flexible, conversational access to campaign data without hiring a BI engineer.