Handing a model a schema and asking for a query produces confident nonsense about columns whose values it has never seen. Distinct values, a random sample, a top-N — those three calls turn guessing into reading, and they cost a fraction of the failed queries they prevent. The registry meta-tools are the other thoughtful piece: past a certain size, listing every tool on every turn is the real cost of a data connection, and searching them instead is the correct fix.
Spice is a data runtime that federates your sources. Its endpoint is per deployed app in Spice Cloud rather than a shared address, and the tools it exposes depend on what you configured.
- Data — list_datasets, table_schema, sql, search
- Sampling — sample_distinct_columns, random_sample, top_n_sample, which let an agent see the shape of a table before querying it
- Runtime — get_readiness, get_current_datetime
- Memory — load and store, plus websearch
- Beyond a certain size the runtime switches to registry meta-tools, so a large tool set is searched rather than listed
An app API key. The runtime itself is open source under Apache 2.0; Spice Cloud is a subscription with usage-based pricing, and there is an enterprise route through AWS Marketplace. The open-source build sends anonymous telemetry by default.
Paste a URL with a key — add the endpoint to your client