MCP tools overview¶
Memcove exposes 12 tools and 2 resources over MCP. Every tool operates strictly within the caller's tenant namespace (resolved from the request headers — see Connect an MCP client).
The descriptions here are the same ones the agent sees at runtime.
The tools¶
| Tool | Purpose | Group |
|---|---|---|
remember_dataset |
Store a table/dataframe/file as a named dataset | Storing |
start_large_upload |
Get a presigned URL to upload a large parquet out-of-band | Storing |
query_memory |
Run a read-only SQL SELECT, get a capped preview | Querying |
derive_dataset |
Save a SELECT result as a new dataset, with lineage | Querying |
recall_dataset |
Read one dataset (preview / schema / stats) without SQL | Reading |
inspect_dataset |
Full metadata: schema, source, tags, lineage | Reading |
list_memory |
List your datasets (optionally filtered by tag) | Reading |
discover_reference_data |
List shared read-only reference schemas/tables | Reading |
export_dataset |
Materialize a dataset/query to a file + presigned URL | Exporting |
stream_dataset |
Get an Arrow Flight ticket to stream a large result out | Streaming |
open_ingest_stream |
Open an Arrow Flight channel to stream a large dataset in | Streaming |
forget_dataset |
Permanently delete a dataset | Deleting |
Resources¶
| Resource URI | Returns |
|---|---|
memcove://{tenant}/{name} |
Metadata for a single dataset |
memcove://{tenant}/_catalog |
List of all datasets for a tenant |
See Resources.
How to read these pages¶
Each tool lists its parameters (name, type, default, and the description the agent sees) and its return shape.
The tenant is implicit
Datasets are always referenced by their bare name in SQL and tool arguments. The tenant is never a parameter — it comes from the request headers and scopes every call automatically. See Connect an MCP client.