rcsb-mcp

An MCP server for interrogating Protein Data Bank structures — discover, inspect, and cross-reference — from LLM clients (Claude Desktop, MCP Inspector, Cursor, etc.). It spans three RCSB APIs:

• Discover — find structures with the Search API (keyword, attribute, sequence, chemistry, 3D shape, motif).
• Inspect — fetch entry / entity / assembly / ligand details and annotations from the Data API.
• Relate — map sequences and positional features across PDB, UniProt, and NCBI with the Sequence Coordinates API.

Tools

Search (search.rcsb.org)

Searching is two steps: build a query with an rcsb_query_* tool, then execute it with rcsb_search_request. The builders are pure — they return a query document (readable JSON plus a digest) and touch no network, so nothing is searched until you call rcsb_search_request. rcsb_query_composer joins two or more documents with AND/OR, which is also how a single search mixes services (e.g. sequence similarity AND an organism filter).

Build

Run

Discover

Both attribute catalogs are generated from the live metadata schemas by scripts/generate_search_attributes.py. To search chemical-component attributes, find the path with rcsb_list_pdb_search_attributes(schema="chemical"), pass chemical_attributes=True to rcsb_query_attribute, and usually set return_type="mol_definition".

Nested attributes. An object can hold many annotations, many binding affinities, many citations. For attributes carrying a nested_group, the query shape selects the semantics: conditions built in one rcsb_query_attribute call, with nothing else in it, must hold on the same record; conditions in separate calls are matched independently against any record. Both are valid and mean different things — type=Kd with value<1 describes one measurement, while an InterPro id and a GO type are necessarily two different annotations.

Counting and faceting are output options on rcsb_search_request, not separate tools: every response includes total_count (the full match count — for "how many ..." run the search with limit=1 and read it), and passing facets returns a breakdown (terms/histogram/date_histogram/range/cardinality) instead of hits. A terms facet also tells you what your hits share, which is how a handful of results becomes a re-searchable value.

Grouping. group_by returns one representative per cluster — seqid_30 … seqid_95 for sequence-identity clusters, or uniprot to collapse by accession — with group_by_ranking choosing the representative. Requires return_type="polymer_entity"; the response reports group_count alongside total_count.

Sorting. sort_by (an attribute path) + sort_direction (asc/desc) replaces the default score ordering (for similarity searches this overrides the similarity-ranked order). Only attributes indexed for sorting work — those exposing exact_match (strings) or equals (numbers/dates) in rcsb_list_pdb_search_attributes; sorting is not available for return_type="mol_definition".

Paging. rcsb_search_request accepts limit (1–100, default 10) and offset (default 0), and each response reports total_count, has_more, and next_offset — call again with the same query document and offset=next_offset. For an explicit "ALL ..." request, all_hits=True returns the complete set in one call (refused above 10,000 hits, and it cannot be combined with offset).

Data (data.rcsb.org/graphql)

There is one tool per Data API GraphQL root field. Each takes a list of IDs (singular lookups = a one-element list) plus an optional fields argument to override the curated default selection with your own GraphQL sub-selection. Unknown IDs are reported under not_found. Discover the paths to put in fields with rcsb_describe_data_object — browse a level, drill into a nested object with into=, or search the schema by keyword with query= + max_depth=. Every path it returns is verified against the live schema, so don't guess field names.

The Search API only returns identifiers, so a search is the first step: batch the returned ids into the matching rcsb_get_* tool to fetch titles, organisms, and other metadata (these tools query the GraphQL endpoint, batching every requested ID into one request). All 16 typed tools are generated from a single registry in queries.py (DATA_OBJECTS), so adding a field or endpoint is a one-line change.

Sequence Coordinates (sequence-coordinates.rcsb.org/graphql)

Maps alignments and positional annotations between sequence reference systems (UNIPROT, NCBI_PROTEIN, NCBI_GENOME, PDB_ENTITY, PDB_INSTANCE). Each tool takes an optional fields argument to override the default selection; use rcsb_describe_seqcoord_object to discover what fields are available.

This is the only RCSB API that cross-references NCBI (RefSeq protein / genome) — the Data API only knows UniProt. So "what NCBI proteins map to a PDB structure?" is answered by rcsb_seqcoord_alignments, not the Data API. PDB query ids must be entity-level (4HHB_1), not a bare entry (4HHB); for a whole entry, query each polymer entity.

Install

# run the published package without installing (recommended for clients)
uvx rcsb-mcp
# or install it
pip install rcsb-mcp

rcsb-mcp is listed in the Official MCP Registry as io.github.rcsb/rcsb-mcp, so registry-aware clients can discover it directly.

For local development, install from the project root instead:

pip install -e .
# or with uv
uv pip install -e .

Run / test

# unit tests (no network)
hatch test          # or: python tests/test_queries.py

# run the server over stdio
python -m rcsb_mcp.server
# or, after install:
rcsb-mcp

# inspect interactively
npx @modelcontextprotocol/inspector python -m rcsb_mcp.server

There are two evaluation suites (evals/): rcsb_pdb_eval.xml, 14 read-only, stable questions measuring how well an LLM can drive these tools to answer real PDB questions, and evals/tool_selection/, a first-tool-call A/B harness for catching routing regressions after a docstring edit. See evals/README.md to run either.

Connect to Claude Desktop

Edit claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "rcsb-mcp": {
      "command": "uvx",
      "args": ["rcsb-mcp"]
    }
  }
}

For a local source checkout, point at the module instead:

{
  "mcpServers": {
    "rcsb-mcp": {
      "command": "python",
      "args": ["-m", "rcsb_mcp.server"],
      "cwd": "/absolute/path/to/rcsb-mcp/src"
    }
  }
}

Restart Claude Desktop. The tools appear under the connectors (plug) icon.

Example prompts

• "Find high-resolution human hemoglobin structures." → rcsb_query_fulltext + rcsb_query_attribute → rcsb_query_composer → rcsb_search_request
• "Human hemoglobin structures better than 2 Å, best resolution first." → same, with sort_by on rcsb_search_request
• "What PDB entries match this protein sequence: MTEY..." → rcsb_query_sequence → rcsb_search_request
• "Find structures containing a ligand like this SMILES / with formula C8H9NO2." → rcsb_query_chemical → rcsb_search_request
• "Which structures have a 3D fold similar to 4HHB?" → rcsb_query_structure → rcsb_search_request
• "Find proteins with a zinc-finger motif." → rcsb_query_seqmotif → rcsb_search_request
• "Structures of proteins with kinase activity / involved in DNA repair / in the mitochondrial membrane." → rcsb_find_go_terms → rcsb_query_attribute on rcsb_polymer_entity_annotation.annotation_lineage.id
• "Structures containing an SH2 domain / immunoglobulin fold." → rcsb_find_interpro_domains → rcsb_query_attribute on rcsb_polymer_entity_annotation.annotation_id
• "Alcohol dehydrogenase structures / any EC 3.4.21 serine protease." → rcsb_find_enzyme_classes → rcsb_query_attribute on rcsb_polymer_entity.rcsb_ec_lineage.id
• "Structures of proteins associated with cystic fibrosis / breast cancer." → rcsb_find_disease_terms → rcsb_query_attribute on rcsb_uniprot_annotation.annotation_lineage.id
• "Structures from mammals / from a particular organism or clade." → rcsb_find_organisms → rcsb_query_attribute on rcsb_entity_source_organism.taxonomy_lineage.id
• "Non-redundant human kinase structures (90% identity clusters)." → rcsb_search_request with group_by="seqid_90", return_type="polymer_entity"
• "How many human X-ray structures are there?" → rcsb_query_attribute → rcsb_search_request with limit=1 (read total_count)
• "Break down ribosome structures by experimental method / by release year." → rcsb_search_request with facets
• "Find structures with the same catalytic-site geometry as residues 162/193/219 of 2MNR." → rcsb_query_strucmotif → rcsb_search_request
• "Find chemical components under 150 Da." → rcsb_list_pdb_search_attributes(schema="chemical") → rcsb_query_attribute with chemical_attributes=True → rcsb_search_request with return_type="mol_definition"
• "Summarize PDB entries 4HHB, 1MBN and 6VXX." → rcsb_get_entries
• "What's the sequence and organism of entity 4HHB_1?" → rcsb_get_polymer_entities
• "Tell me about the ligand HEM." → rcsb_get_chem_comps
• "What's the composition of the 4HHB biological assembly?" → rcsb_get_assemblies
• "Which PDB entries does P69905 map to?" → rcsb_get_uniprot
• "Which PDB entities align to UniProt P69905, and over what ranges?" → rcsb_seqcoord_alignments
• "What NCBI proteins map to 4HHB?" → rcsb_seqcoord_alignments per entity (4HHB_1, 4HHB_2), to_ref=NCBI_PROTEIN
• "Show UniProt features mapped onto PDB entity 4HHB_1." → rcsb_seqcoord_annotations
• "Pull a field the compact defaults don't include." → rcsb_describe_data_object to find the path, then the matching rcsb_get_* tool with fields=

Report output (RCSB_MCP_REPORT_BASE_URL)

rcsb_render_report returns a self-contained link, not the markup — so the agent never has to reproduce the ~20 KB document (the single most expensive step of a report turn). The whole report is gzip+base64url-packed into the URL, so the server stores nothing and any replica renders any link on demand.

Set RCSB_MCP_REPORT_BASE_URL to the origin that serves this MCP (e.g. https://rcsb-mcp.rcsb.org) and the tool returns { url: "<base>/r?d=<packed report>", html: null }. The agent hands the user that link; opening it hits the stateless render endpoint:

GET /r?d=<gzip+base64url of the report JSON>   →   text/html

The endpoint decodes, validates against the report schema, and renders with the fixed template. It is hardened for a public route: the d token and its decompressed size are both capped (a gzip bomb is refused before it expands), the page carries Content-Security-Policy: default-src 'none' + noindex, and every value is escaped by the template — a crafted link can only ever produce an escaped report.

Fallback. When RCSB_MCP_REPORT_BASE_URL is unset (e.g. local stdio dev with no reachable endpoint) or a report is too large to pack into a URL, the tool returns html instead of url, and the rcsb_search_assistant prompt tells the model to write it to a .html file. Reports compress to ~1 KB even at 50 rows, so the size fallback is rare.

Notes

• Search endpoint: https://search.rcsb.org/rcsbsearch/v2/query (POST, JSON body).
• Data endpoint: https://data.rcsb.org/graphql (POST, GraphQL). It returns HTTP 200 even for query errors, reporting them in an errors array.
• Sequence Coordinates endpoint: https://sequence-coordinates.rcsb.org/graphql (POST, GraphQL; same HTTP-200-with-errors behavior).
• The rcsb_find_* resolvers map free text to ontology ids via EBI services — the non-RCSB dependencies: GO via QuickGO (.../QuickGO/services/ontology/go/search), InterPro and Pfam via EBI Search (.../ebisearch/ws/rest/interpro7), EC via EBI Search/IntEnz (.../ebisearch/ws/rest/intenz), disease via OLS/MONDO (.../ols4/api/search?ontology=mondo), and organisms via UniProt (.../rest.uniprot.org/taxonomy/search). The resolved ids then drive RCSB annotation searches (rcsb_polymer_entity_annotation.*, rcsb_polymer_entity.rcsb_ec_lineage.id, rcsb_uniprot_annotation.annotation_lineage.id).
• No API key required; the APIs are public. Be considerate with request volume.
• Every outbound request sends User-Agent: rcsb-mcp/<version> (https://github.com/rcsb/rcsb-mcp), <version> being the installed package's (pyproject.toml's) version. Set RCSB_MCP_USER_AGENT to override it (printable ASCII; a {version} in it is filled in the same way); the Helm chart does, so the hosted service's traffic is distinguishable from local installs in upstream logs.
• A full list of searchable attributes for rcsb_query_attribute is in the Search API attribute reference; the Data API schema is documented at data.rcsb.org/index.html#gql-api.

Prompt

The server exposes one MCP prompt, rcsb_search_assistant ("RCSB PDB search assistant") — the search requirements plus the HTML-report output format. Because it is served over the protocol's prompts capability, any MCP client can list and invoke it (Claude Desktop surfaces server prompts in the + / prompt menu); there's nothing to copy-paste. The text lives in src/rcsb_mcp/prompts/rcsb_search_assistant.md and ships with the package.

Invoke it when you want answers formatted as a PDB report. It is not required for the tools to work: every routing rule, gotcha and cross-reference lives on the tool descriptions themselves, which the protocol always delivers. A server instructions block and a second rcsb_mcp_guide prompt both used to carry that guidance and were retired — instructions because clients truncate or drop it, the prompt because it is opt-in and may never be loaded. prompts/rcsb_mcp_guide.md is kept on disk, unserved, as a source to rescue prose from.