A Model Context Protocol server for ClickHouse. It lets an MCP-capable client (Claude Desktop, Claude Code, etc.) explore schemas, run analytical queries, and manage the database — with behaviour controlled entirely by flags.
The security model is statement-aware: every SQL statement is classified as read, write, or destructive, and gated against the current access mode. Read-only mode additionally runs queries under ClickHouse's own readonly=1 setting.
- Exploration & monitoring — databases, tables, columns,
SHOW CREATE, table stats (parts/rows/bytes), running queries, server metrics, cluster topology. - Read queries — a
querytool that only accepts read statements, capped atCLICKHOUSE_MAX_ROWS. - Management — an
executetool for INSERT/CREATE/ALTER (read-write) and DROP/TRUNCATE/DELETE (admin), each gated by classification. - Access modes —
read-only→read-write→admin, layered so a mode never exposes statements above its level. - Security flags — database allowlist, protected databases, destructive gating, row cap, dry-run, and JSON audit logging (see below).
| Concern | Flag | Default | Effect |
|---|---|---|---|
| What can the server do? | CLICKHOUSE_MODE |
read-only |
read-only exposes read tools only (and refuses non-SELECT in query); read-write adds execute for writes; admin allows destructive statements. |
| Which databases are in scope? | CLICKHOUSE_DATABASE_ALLOWLIST |
(all) | When set, a statement that touches any other database is refused: the databases are read off the SQL itself (qualified names, USE, joins, subqueries, INSERT … SELECT, RENAME, …) and unqualified names count as the connection's default database (CLICKHOUSE_DATABASE). A statement the server cannot analyse is refused. Best-effort — back it with ClickHouse grants. |
| Which databases are read-only forever? | CLICKHOUSE_PROTECTED_DATABASES |
system,information_schema |
Readable, never mutable. |
| Can it run destructive SQL? | CLICKHOUSE_ALLOW_DELETE |
false |
DROP/TRUNCATE/DELETE/RENAME/EXCHANGE/data-losing ALTER (DROP PARTITION/COLUMN/INDEX, DETACH/CLEAR/REPLACE PARTITION, RENAME COLUMN, MOVE … TO TABLE, MODIFY TTL, type-changing MODIFY COLUMN, mutations), CREATE OR REPLACE, `OPTIMIZE … DEDUPLICATE |
| Result size cap | CLICKHOUSE_MAX_ROWS |
1000 |
Hard cap on rows returned to the model. |
| Preview without executing | CLICKHOUSE_DRY_RUN |
false |
Write/destructive statements validate + log intent, then return. |
| Audit trail | CLICKHOUSE_AUDIT_LOG |
true |
Emits a JSON line to stderr per guarded operation. |
| Interactive confirmation | (automatic) | — | Destructive statements prompt the human to approve via MCP elicitation before running. The prompt shows the tool, the exact arguments (the full SQL) and a fingerprint; for DROP / TRUNCATE / DELETE / RENAME / EXCHANGE / ALTER / CREATE OR REPLACE / OPTIMIZE on a table, view, dictionary or database named in the statement, that object's version is pinned across the prompt and the statement is refused if it changed; DROP TABLE / DROP DATABASE also ask the human to type the name; a prompt nobody answers is a refusal. |
| Require a human | CLICKHOUSE_REQUIRE_HUMAN_APPROVAL |
false |
When true, destructive statements are refused if the client cannot elicit (instead of falling back to the CLICKHOUSE_ALLOW_DELETE gate alone). |
Statement classification lives in src/sql.ts and is fail-safe: ALTER … DELETE/UPDATE and the data-losing ALTER commands, CREATE OR REPLACE and OPTIMIZE … DEDUPLICATE|CLEANUP count as destructive, and anything unparseable is treated as destructive.
Read (read-only+): list_databases, list_tables, describe_table, show_create_table, table_stats, running_queries, server_metrics, cluster_info, query
Write/Admin (read-write+): execute — runs a single statement after classifying it; writes need read-write mode, destructive statements need admin mode + CLICKHOUSE_ALLOW_DELETE.
Published on npm as @dockndevai/mcp-clickhouse. No clone or build needed — your MCP client runs it on demand with npx. Start in read-only mode; see .env.example for every variable and docs/CLIENTS.md for the full per-client guide.
Claude Code (CLI)
claude mcp add clickhouse -e CLICKHOUSE_URL="http://localhost:8123" -e CLICKHOUSE_USER="default" -e CLICKHOUSE_MODE="read-only" -- npx -y @dockndevai/mcp-clickhouse
Claude Desktop · Cursor · Windsurf — same block in claude_desktop_config.json, .cursor/mcp.json, or ~/.codeium/windsurf/mcp_config.json:
{
"mcpServers": {
"clickhouse": {
"command": "npx",
"args": [
"-y",
"@dockndevai/mcp-clickhouse"
],
"env": {
"CLICKHOUSE_URL": "http://localhost:8123",
"CLICKHOUSE_USER": "default",
"CLICKHOUSE_MODE": "read-only"
}
}
}
}OpenAI Codex CLI — in ~/.codex/config.toml:
[mcp_servers.clickhouse] command = "npx" args = ["-y", "@dockndevai/mcp-clickhouse"] env = { CLICKHOUSE_URL = "http://localhost:8123", CLICKHOUSE_USER = "default", CLICKHOUSE_MODE = "read-only" }
VS Code (GitHub Copilot, Agent mode) — in .vscode/mcp.json:
{
"servers": {
"clickhouse": {
"type": "stdio",
"command": "npx",
"args": [
"-y",
"@dockndevai/mcp-clickhouse"
],
"env": {
"CLICKHOUSE_URL": "http://localhost:8123",
"CLICKHOUSE_USER": "default",
"CLICKHOUSE_MODE": "read-only"
}
}
}
}- "What are the biggest tables in the
analyticsdatabase?" - "Show me the schema for
eventsand run a query for daily counts this week." - "Which queries are currently running and using the most memory?"
Prefer the published package above. To run from a clone:
npm install
npm run build
node dist/index.js # with the environment variables setnpm run dev npm test # SQL classification + security policy (30 tests) npm run typecheck
This server ships a server.json for the official MCP registry and an mcpName for npm ownership validation. See PUBLISHING.md for publishing to npm and listing on the MCP registry, Smithery, Glama, Cursor, and PulseMCP.
MIT