cb-analytics-mcp

A standalone MCP (Model Context Protocol) server plus admin GUI for Couchbase Enterprise Analytics and Couchbase Capella Analytics. Connects to Claude (claude.ai, Claude Desktop, Claude Code) and exposes the full Analytics REST surface as 55 well-typed, audited tools.

© 2026 Chris Ahrendt — MIT licensed

What you get

• 55 MCP tools spanning query execution, schema introspection, ingestion control, link management, UDF libraries, RBAC, cluster admin, and Capella cloud management.

• Paginated queries for large result sets, with TTL caching for repeated reads, an explain_query tool for plan investigation, and a configurable soft cap on full-result queries (default 1000 rows; truncated responses include truncated: true and full_row_count).

• Rate limiting per API key, per tool category (10 query/sec, 60 read/sec, 1 write/sec by default) so a runaway client can't melt your cluster.

• Multi-cluster support out of the box (one process, many Couchbase clusters).

• Bundled web GUI at :8080 with login, SQL++ editor, paginated audit-log viewer with date / tool / failure filters, per-cluster Test Connection buttons, config explorer, and live log tail (WebSocket-streamed, with HTMX polling fallback).

• CLI for testing tools without a running server (cb-analytics-mcp tools call <name> --offline) or against a remote one (--remote URL).

• Structured logging + audit trail + Prometheus metrics + OpenTelemetry tracing — all opt-in but on by default for the basics. Audit log rotates at 10 MB by default and keeps 5 generations.

• Hardened HTTP client with retry, exponential backoff, async-native circuit breaker, and consistent error mapping.

• Strict scans — every release passes pyflakes, ruff, mypy strict, bandit, and pytest with 94 %+ coverage.

Related MCP server: Datris MCP Server

Quickstart

# 1. Clone and install
git clone https://github.com/celticht32/cb-analytics-mcp
cd cb-analytics-mcp
make dev

# 2. Configure
cp .env.example .env
$EDITOR .env          # set MCP_API_KEY, CB_ANALYTICS_HOST, etc.

# 3. Run
make run              # starts MCP server on :8000 + GUI on :8080

Then open http://localhost:8080 in a browser, or point Claude at http://localhost:8000/mcp with the bearer token you set as MCP_API_KEY.

For reproducible builds in CI / Docker, install from the locked requirements.txt instead of pyproject.toml:

pip install -r requirements.txt
pip install --no-deps -e .

Contributors can opt in to pre-commit hooks (ruff, mypy, bandit run on every commit):

pip install pre-commit && pre-commit install

Documentation

• Installation

• Configuration

• Connecting Claude.ai / Desktop / Code

• GUI walkthrough

• Tool reference (all 55 tools)

• Observability — logs, metrics, traces, audit

• Security model

• Deployment (Docker, Compose, TLS)

• Architecture

Project layout

cb-analytics-mcp/
├── src/cb_analytics_mcp/
│   ├── auth.py              # bearer-token verifier
│   ├── config.py            # env-driven config loader
│   ├── pool.py              # connection pool (clusters + Capella)
│   ├── server.py            # FastMCP wire-up
│   ├── __main__.py          # entrypoint (server / GUI / --check)
│   ├── couchbase/           # internal HTTP client (no SDK dep)
│   ├── observability/       # logging, audit, metrics, tracing
│   ├── tools/               # 10 modules × 55 MCP tools
│   └── gui/                 # FastAPI + HTMX admin GUI
├── tests/unit/              # 411 unit tests, mocked I/O
├── docs/                    # MkDocs-ready Markdown
├── .claude/skills/          # 8 Claude Code skills
├── scripts/                 # deep_scan.sh, make_screenshots.py
└── pyproject.toml

Make targets

Quality bar

Every release must pass make scan. Current numbers:

Future possibilities

Things that aren't built yet but would be obvious next steps. Issues and PRs welcome on any of these:

• Query history in the GUI. The query editor loses your statement on page reload. A small dropdown of recent queries (in-memory only, scoped to the current GUI session) would be ergonomic. The audit log already has every executed statement; this is just surfacing it in the editor.

• DDL helper tools. create_dataverse, create_dataset, drop_dataset, create_index, drop_index as typed tools instead of relying on Claude to compose raw SQL++. Safer, less ambiguous.

• Bucket / scope / collection awareness. Today's schema tools work at dataverse level, but real Couchbase deployments use buckets → scopes → collections. list_buckets, list_scopes(bucket), list_collections(...) would help Claude navigate clusters it hasn't seen before.

• Integration tests against a real Couchbase. The 519 unit tests use mocks; a small tests/integration/ suite that spins up couchbase/server:enterprise-7.6 in a Docker container would catch things mocks can't (real timeouts, real auth flows, edge cases in the Analytics REST surface).

• Per-team / per-project API keys. Today the MCP server accepts a single bearer token. Larger deployments would want per-team tokens with their own audit attribution and RBAC scope.

• Tracing across the Couchbase REST boundary. OpenTelemetry currently stops at the MCP boundary. Instrumenting the outgoing httpx requests would give full traces from Claude → MCP → Couchbase.

• Distributed rate limits / pagination state. The current implementations are in-process. Scaling out horizontally would need a shared store (Redis or similar).

Acknowledgements

This project speaks to Couchbase through its public REST API. It is not an official Couchbase product. "Couchbase" and "Couchbase Capella" are trademarks of Couchbase, Inc.

License

MIT — see LICENSE.