Labsco
MCP SERVER

Run SQL++ against a Couchbase cluster, read documents by key or by path, and see which queries are hurting.

NoSQL, Graph & Key-Value StoresVerified
Summary
Read-only out of the box, and unusually strong on the question of which queries are costing you.

The query-monitoring tools are the part you will not find on most database servers: seven separate reads of the completed_requests catalog that surface unselective queries, missing covering indexes and primary-index scans, with the Index Advisor available on any statement you paste in. Writes exist across documents, collections and indexes, but they stay switched off until you set the read-only flag to false, and individual tools can be disabled or made to ask first.

What it is

A Couchbase server spanning three surfaces: cluster and collection structure, key-value document access, and SQL++ with the query monitoring catalog behind it.

What you get
  • SQL++ execution scoped to a bucket and scope, so queries name collections directly, plus an EXPLAIN plan with its findings evaluated
  • Structure discovery from the top down: buckets, scopes, collections, and an inferred schema for a collection
  • Documents by ID, or just the parts you need — subdocument lookup fetches specific fields, existence checks and array or object counts without pulling the whole document
  • Writes when you enable them: insert, replace, upsert and delete by ID, and path-level mutation covering array operations and counters
  • Collection and index management — create and delete scopes and collections, create, build and drop GSI indexes
  • Index work backed by evidence: list the indexes with their definitions, and get Index Advisor recommendations for a specific SQL++ query
  • Seven readings of the completed_requests catalog: longest running, most frequent, largest responses, largest result counts, queries on a primary index, queries without a covering index, and queries that are not selective
  • Operational checks: cluster health and running services, a credential test against cluster and bucket, cached SDK connection diagnostics, and a configuration read that works before any connection exists
Requirements

A Couchbase cluster and credentials: CB_CONNECTION_STRING plus either CB_USERNAME and CB_PASSWORD, or the mTLS pair CB_CLIENT_CERT_PATH and CB_CLIENT_KEY_PATH. Self-signed clusters also need CB_CA_CERT_PATH; Capella does not. Installed from PyPI as couchbase-mcp-server, stdio by default, with http and sse available through CB_MCP_TRANSPORT plus CB_MCP_HOST and CB_MCP_PORT. CB_MCP_READ_ONLY_MODE is true by default — every write tool stays unloaded and data-modifying SQL++ is blocked until you turn it off. CB_MCP_DISABLED_TOOLS and CB_MCP_CONFIRMATION_REQUIRED_TOOLS narrow the surface further, the latter needing a client that supports elicitation.

Setup effort

One command plus a key — uvx couchbase-mcp-server, then supply credentials