Spelunk

A multi-source DuckDB query + transformation-pipeline MCP server. Point it at files (CSV/Parquet/JSON/Excel) and databases (SQLite/PostgreSQL/MySQL), and an agent like Claude Code can query across all of them — and build step-by-step pipelines — through one DuckDB engine.

Architecture in one line

A single DuckDB session is both the query engine and the workspace: every source is ATTACHed (databases) or scanned (files) into one connection, so one query can join a Parquet file to a Postgres table to a result you built two steps ago — all in DuckDB SQL.

Tools

query(sql, name, flow?)     # run a read-only SELECT over sources + results; store the full
                            #   result as table `name` for immediate reuse. The ONE tool for
                            #   looking and building — every result is named and chainable.
profile(sql, flow?)         # per-column stats (null_rate, min/max/mean/std, percentiles, top/freq)
export(target, fmt, path)   # write a saved result name OR a full SELECT to csv/json/parquet
catalog(flow?)              # list flows, or the results in one flow
drop(name?, flow?)          # drop one result, or a whole flow
lineage(name?, flow?)       # provenance DAG: the SQL + deps that built a result (or a whole flow)
replay(flow?, into?)        # rebuild a flow from its recorded SQL, in dependency order

Discovery resources: db://tables (queryable source objects) and db://{table} (columns, PK, sample, row count). A flow is an isolated result namespace (a DuckDB schema); give each concurrent line of analysis its own flow.

Install & run

The fastest path is uvx — no clone, no venv. It fetches Spelunk into an ephemeral environment and runs the spelunk command. The package is published as spelunk-mcp (the command is spelunk), so pass it via --from:

uvx --from spelunk-mcp spelunk \
  --source sales=./data/sales.parquet \
  --source sqlite:///path/to/app.db \
  --session-dir .spelunk_session          # omit for an ephemeral (non-durable) workspace

Want the latest commit instead of the released version? Point uvx straight at the repo:

uvx --from git+https://github.com/shaunak2512/spelunk spelunk --source sales=./data/sales.parquet

Either way, wire it into Claude Code with a .mcp.json:

{
  "mcpServers": {
    "spelunk": {
      "command": "uvx",
      "args": ["--from", "spelunk-mcp", "spelunk", "--source", "sales=./data/sales.parquet"]
    }
  }
}

Prefer a local checkout? python -m spelunk.mcp.server --source ... is equivalent to the spelunk command.

Sources auto-detect by extension/scheme; prefix with name= to set the catalog/view name. Optional resource guards: --memory-limit 4GB, --temp-dir <dir>, --max-temp-size 50GB. DuckDB is out-of-core, so a source larger than RAM is the normal case — scans read on demand and buffering operators spill to the temp directory.

Dev

uv sync --extra dev
uv run --extra dev python -m pytest -q
uv run --extra dev ruff check spelunk/