datapipelines.co
Menu

Generated from the server's own catalogue

The MCP tools: everything an agent can do against your databases

41 tools over Streamable HTTP, 16 of them writes, every one gated by the scope its key must hold. The list below is rendered from the same catalogue the server registers at startup, so it cannot drift from what your agent sees in tools/list. What is not here is as deliberate as what is: there is no release tool, no promote tool, and no tool that returns a credential.

Add it to Claude Code Every tool's schema

By family

Read what it may, write only drafts

Names are <family>_<verb>. Reads need read; anything that runs SQL against your data needs execute; anything that creates or changes a draft, registers a lake table or probes a SELECT needs author. Previewing rows and probing SQL are author rather than read, because a schema is metadata and a row is your customer's; catalogue statistics stay read — a count the engine already stored is shape, not data.

The three semantics_* tools record what an agent learned about your data — a unit, a time zone, a grain, what a code means — so the next session reads it beside the columns. The learned semantic layer →

Pipelines — author, run, inspect

  • pipelines_list read
  • pipelines_get read
  • pipelines_execute execute writes
  • pipelines_execute_node author writes
  • pipelines_create author writes
  • pipelines_update author writes
  • pipelines_run_checks execute writes

Templates — the SQL an agent writes

  • templates_list read
  • templates_get read
  • templates_used_by read
  • templates_create author writes
  • templates_update author writes
  • templates_render author
  • templates_purge_draft author writes

Datasources — what the agent may see

  • datasources_list read
  • datasources_get read
  • datasources_test author
  • datasources_get_schemas author
  • datasources_get_tables author
  • datasources_get_columns author
  • datasources_get_table_stats read
  • datasources_preview_rows author

dp-lake — tables over your bucket

  • lake_tables_register author writes
  • lake_tables_import author writes
  • lake_tables_unregister author writes

The SQL probe

  • sql_probe author

Executions — results and cancellation

  • executions_list read
  • executions_get read
  • executions_get_result read
  • executions_cancel execute writes

Published endpoints

  • endpoints_create author writes
  • endpoints_list read
  • endpoints_get read
  • endpoints_delete author writes

Calculators

  • calculators_list read
  • calculators_get read

docs

  • docs_list read
  • docs_get read

semantics

  • semantics_record author writes
  • semantics_list read
  • semantics_retire author writes

The two probes

Statistics before the predicate, a plan before the timeout

datasources_get_table_stats

Row estimates, indexes and per-column distinct counts, null fractions and ranges — from the engine's catalogue, never a scan. An agent reads them before it writes a WHERE, and suggests the index that would help when none does.

sql_probe

One classified SELECT, named typed parameters, a row cap and a time box; the EXPLAIN is captured before the run so the plan survives a timeout. On dp-lake it reports how many partitions the predicate touched.

What is missing on purpose

No pipelines_release, no promote, no datasources_create with credentials. An agent leaves a draft; a person releases; an operator registers datasources. The absence is the design.

Give your agent the skill, not just the tools

The server also ships a skill — the rules an agent follows to author a pipeline well: filter lake tables on their partition column, pre-aggregate before staging, never divide decimals in the scratch engine, leave the release to a person. It is served at /skill.md, exposed as an MCP resource, and packaged as a Claude Code plugin.

Add the server and the skill How access is scoped