Altinity's agent skills for ClickHouse® development and DBA tasks
This repository also configures its own agents. See what altinity-skills tells them →
Skill Claude CodeCodex
Analyzes the ClickHouse mark cache, uncompressed cache, query cache and compiled expression cache, their hit ratios and their RAM footprint. Use when cache hit ratios are poor, caches take too much RAM, mark counts look excessive, or cache sizing needs tuning.
Skill Claude CodeCodex
Establishes how to reach ClickHouse (MCP tool or clickhouse-client), the cluster name for clusterAllReplicas('{cluster}', ...) query packs, the default 24-hour time window, and the shared report header. Use first, before any other altinity-expert-clickhouse skill.
Skill Claude CodeCodex
Diagnoses ClickHouse external dictionaries - load status, memory usage, layout choice, reload staleness, source connectivity and lookup performance. Use for dictionary load failures, FAILED or LOADINGFAILED status, stale dictionary data, high dictionary RAM and slow dictGet lookups.
Skill Claude CodeCodex
Diagnose and resolve ClickHouse grant and authentication errors, especially after upgrades. Use when queries fail with ACCESSDENIED/NOTENOUGHPRIVILEGES, AUTHENTICATIONFAILED/WRONGPASSWORD/REQUIREDPASSWORD, or ON CLUSTER privilege errors; when system. or INFORMATIONSCHEMA access is denied; or when grant behavior…
Skill Claude CodeCodex
Analyze whether ClickHouse indexes (PRIMARY KEY, ORDER BY, skipping indexes, projections) are effective for the actual query patterns. Use when investigating index effectiveness, ORDER BY key design, query-to-index alignment, or when queries scan far more data than expected.
Skill Claude CodeCodex
Diagnoses ClickHouse INSERT performance, batch sizing, part creation patterns and materialized-view overhead on inserts. Use for slow inserts, failed inserts, micro-batching, high part creation rate, Buffer table flushes and data pipeline issues.
Skill Claude CodeCodex
Diagnose ClickHouse Kafka engine health, stuck consumers, thread pool capacity, consumer lag, slow materialized views, and Kafka rack-awareness or cross-AZ traffic. Use for Kafka lag, consumer errors, rebalances, thread starvation, or unexpected cross-AZ Kafka cost.
Skill Claude CodeCodex
Checks ClickHouse system log tables for freshness, TTL, disk footprint, upgrade leftovers and crash records. Use when system tables eat the disk, querylog or partlog has no recent data, retention for system.log needs review, or the server appears to have crashed.
Skill Claude CodeCodex
Diagnoses ClickHouse RAM usage, memory limits, OOM errors and the allocation breakdown across queries, caches, dictionaries and primary keys. Use when MEMORYLIMITEXCEEDED appears, the server is killed by the OOM killer, memory grows steadily, or GROUP BY and JOIN queries run out of memory.
Skill Claude CodeCodex
Diagnose ClickHouse merge performance, part backlog, merge memory, and "too many parts" errors. Use when merges look stopped or slow, parts pile up, TOOMANYPARTS or partstothrowinsert is hit, or merges fail with MEMORYLIMITEXCEEDED.
Skill Claude CodeCodex
Resource saturation from system.metrics, system.events and system.asynchronousmetrics - memory, load average, connections, running queries, background pools, parts per partition, block device queues - plus short history. Use when the server feels slow or overloaded, queries queue or get rejected, or load and…
Skill Claude CodeCodex
Diagnoses ClickHouse mutations - ALTER UPDATE, ALTER DELETE, MATERIALIZE and column changes. Use for stuck mutations, mutations that never finish, latestfailreason errors, growing mutation queues and slow ALTER operations.
Skill Claude CodeCodex
Fast ClickHouse server health snapshot (object counts, memory, disks, replication summary, system log TTLs, error rates, background pools, detached parts) that routes to the specialist skills. Use as the entry point for "is the server healthy" or when the problem area is not yet known.
Skill Claude CodeCodex
Diagnoses part lifecycle problems from system.partlog - part creation, merges, mutations, downloads, removals and moves. Use when there are too many parts or micro-batch inserts, merge backlog or slow merges, mutation storms from ALTER UPDATE or DELETE, unusual DownloadPart replication churn, unexpected RemovePart…
Skill Claude CodeCodex
Diagnose ClickHouse replication health, Keeper connectivity, read-only replicas, replica lag, replication queue backlog and slow fetches. Use for replication lag, read-only replica problems, growing replication queues, failed fetches, or Keeper/ZooKeeper session and latency errors.
Skill Claude CodeCodex
Diagnoses ClickHouse SELECT query performance, query patterns, slow queries, failures and optimization opportunities. Use for query latency, timeouts, high CPU queries, queries reading too much data and repeated expensive query patterns.
Skill Claude CodeCodex
Diagnoses ClickHouse table design - partitioning, ORDER BY and primary key, column count, Nullable usage, identifier length and materialized view structure. Use for schema anti-patterns, too many partitions, oversized partitions, bad sort keys and materialized view design problems.
Skill Claude CodeCodex
Read-only ClickHouse security audit of users, roles, grants, row policies, settings profiles, quotas, named collections, table functions and engines, executable UDFs, TLS and network exposure, Keeper and interserver security, audit logging, password hash hygiene, and SQL SECURITY DEFINER. Use when assessing security…
Skill Claude CodeCodex
Diagnoses ClickHouse disk usage, compression efficiency, part sizes, storage policies and IO throughput. Use when disks fill up, data grows faster than expected, compression looks poor, IO or merges are slow, or detached and tiny parts accumulate.
Skill Claude CodeCodex
Profile a ClickHouse cluster via MCP and emit a per-cluster "analyst" Skill the user can save in claude.ai. Activate when the user asks to "profile this ClickHouse", "generate an analyst skill", "build a schema guide", "map the data in this cluster", or regenerate an existing cluster-analyst Skill after schema…
At most 3 mods per repository are shown here, and a mod shipped inside a plugin is left to that plugin's page — the rest are on their repository pages: