Summary

Schema first, then query — which is what stops a model inventing field names.

Both halves follow the same shape: fetch the schema, then run the query. That matters here because the search index and the PDBe-KB graph both have vocabularies nobody memorises, and a guessed field name returns nothing rather than an error. The graph schema tools work with no database of your own; only Cypher execution needs a local install.

What it is

A set of servers over PDBe. One generates its tools from the PDBe REST API for core entry data; a second exposes PDBe's Solr search; a third inspects the PDBe-KB graph schema and, against a graph database of your own, runs read-only Cypher.

What you get

Core PDBe entry data — entries, assemblies, molecules, ligands, publications and validation — through tools generated from the API specification
The full Solr search schema, so field names and types come from the index rather than from memory (get_pdbe_search_schema)
Structure search with field selection, filter queries, facets, grouping, sorting and paging (run_pdbe_search_query)
The graph schema without a database of your own: node types, relationship types, the patterns defined for a given label, and worked example queries (pdbe_graph_nodes, pdbe_graph_edges, pdbe_graph_node_relationships, pdbe_graph_example_queries)
Read-only Cypher against a PDBe-KB graph you host yourself — only MATCH and OPTIONAL MATCH are accepted, and writes are refused (pdbe_run_cypher_query)
An experimental compact output format for API and Cypher responses, which falls back to JSON if encoding fails

Requirements

Python and uv; the package runs straight from PyPI with no install step. The API and search servers need nothing else. The Cypher tool needs a PDBe-KB graph database you set up and credential yourself — PDBe does not host a public one, so most people should start with the API and search servers.

Setup effort

One command — uvx pdbe-mcp-server

PDBe MCP Servers