BigQuery

BigQuery MCP gives Claude direct read access to your Google BigQuery warehouse. You authenticate via a service account JSON file, point it at your project, and Claude can then list datasets, inspect schemas, run queries, and estimate costs before execution. The schema discovery is genuinely useful: ask Claude to summarise your table structure and it'll pull column names, types, and descriptions without you writing a single SELECT statement. I'd reach for this when I'm exploring unfamiliar datasets or prototyping analytics queries in conversation. The dry-run cost estimation is the standout feature. BigQuery can get expensive fast, and being able to ask 'how much would this query cost?' before running it saves both money and surprises. Parameterised queries work as expected: you can template SQL with placeholders and Claude will fill them in based on context. The main quirk is that it's read-only by design. You can't CREATE, INSERT, or DELETE anything, which is sensible for an MCP server but means you'll still need the BigQuery console or a proper client for write operations. Authentication requires a service account with BigQuery Data Viewer and Job User roles at minimum. If your organisation locks down service account creation, you'll need to negotiate with whoever controls IAM. Performance depends entirely on your dataset size and query complexity. Claude doesn't cache results, so repeated queries hit BigQuery every time. For large tables, you'll want to guide Claude toward filtered queries or sampled data to avoid slow responses. The server itself is lightweight and doesn't add noticeable overhead. This is for people who already use BigQuery and want conversational access to their data. If you're not running a warehouse, this does nothing for you. If you are, it's a faster way to explore schemas and prototype queries than switching between Claude and the BigQuery console. The cost estimation alone justifies the setup time.