UniProt Link
berntpopp/uniprot-link
MCP server para pesquisa estruturada de proteínas via UniProt SPARQL
Ver no GitHubServidor MCP que integra a base de dados UniProt SPARQL para pesquisa de proteínas. Oferece 15 ferramentas tipadas para consultar estruturas de proteínas, sequências, variantes, doenças e termos GO, além de um executor SPARQL com 126 queries curadas e suporte a federação.
Instalação
npx convez-ai berntpopp/uniprot-link1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Pesquisa em biologia molecular e genômica computacional
- Análise de estruturas e variantes de proteínas
- Automação de consultas bioinformáticas com LLM
- Integração de dados UniProt em pipelines de pesquisa
Veredito

Ferramenta especializada para pesquisadores em bioinformática que precisam consultar proteínas estruturadas via SPARQL. Arquitetura sólida com agentic affordances e escape hatch para queries customizadas, mas requer domínio técnico em biologia molecular e SPARQL.
README
A tradução para o português deste README ainda está na fila.
uniprot-link
An MCP (Model Context Protocol) + REST server that grounds protein research in the
UniProt SPARQL endpoint (https://sparql.uniprot.org/sparql). It wraps a
~232-billion-triple, QLever-backed SPARQL 1.1 service behind intent-named,
token-economical tools — and ships a safe raw-SPARQL escape hatch plus UniProt's
126 curated example queries so an LLM can learn and write its own queries.
Part of the *-link family of biomedical MCP servers (gnomad-link, gtex-link,
pubtator-link, genereviews-link, …) and follows their stack and conventions.
Features
- 15 MCP tools across discovery, raw SPARQL, the curated example catalog, proteins (UniProtKB), and taxonomy.
- Typed protein tools —
get_protein, sequence, features (with FALDO coordinates), natural variants, diseases, GO terms, cross-references, id mapping. search_sparql_query— execute any SPARQL 1.1 query (SELECT/ASK/CONSTRUCT/ DESCRIBE, withSERVICEfederation); auto-LIMIT on unbounded SELECTs; JSON/XML/ CSV/TSV/Turtle/RDF-XML/N-Triples output.- Example catalog —
search_example_queries/get_example_queryexpose UniProt's 126 curated, executable queries (backed by thesparql-examplesgraph). - Agentic affordances — every response carries
_meta.next_commands(ready-to-run{tool, arguments}steps), a structured error taxonomy, and auniprot://capabilitiesdiscovery resource. - Two transports — unified (REST + MCP/HTTP) and HTTP-only. Streamable HTTP only.
Quick start
make install # uv sync --group dev
make dev # unified server: REST on / and MCP on /mcp (port 8000)
uv run uniprot-link serve --help # CLI: serve / config / health / version
make ci-local # format + lint + loc + typecheck + tests
CLI
The uniprot-link console script is the single entry point (Streamable HTTP only):
uniprot-link serve --transport unified --host 127.0.0.1 --port 8000 # REST + MCP/HTTP
uniprot-link serve --transport http # REST only
uniprot-link config --validate # show + validate config
uniprot-link health --url http://127.0.0.1:8000 # probe /health
uniprot-link version # print version
Connect an MCP client
Point a Streamable-HTTP MCP client at the /mcp endpoint of a running server, e.g.:
claude mcp add --transport http uniprot-link --scope user http://127.0.0.1:8000/mcp
Tool catalog
| Tool | Purpose |
|---|---|
get_server_capabilities | Tools, 21 named graphs, prefixes, formats, workflows, limits |
search_sparql_query | Execute any SPARQL 1.1 query (the power tool / federation) |
search_example_queries | Search 126 curated example queries |
get_example_query | Full text + metadata of one example |
find_proteins | Search UniProtKB by gene_symbol / organism / keyword / EC / mnemonic |
get_protein | Core entry summary by accession |
get_protein_sequence | Canonical + isoform sequences |
get_protein_features | Sequence features with FALDO coordinates |
get_protein_variants | Natural-variant annotations |
get_protein_diseases | Disease annotations |
get_protein_cross_references | Cross-references grouped by database |
get_protein_go_terms | GO annotations by aspect |
resolve_identifiers | Resolve an accession to external DB ids |
get_taxon | Resolve a taxon by id or name |
Tool names are unprefixed (the GeneFoundry Tool-Naming Standard v1): the server reports
serverInfo.name = "uniprot-link"and its canonical gateway namespace token isuniprot. When federated behind thegenefoundry-routergateway, tools surface asuniprot_<tool>(e.g.uniprot_find_proteins); standalone MCP clients already namespace them asmcp__uniprot-link__<tool>.
Configuration
Environment variables (prefix UNIPROT_LINK_, nested with __):
| Variable | Default | Meaning |
|---|---|---|
UNIPROT_LINK_SPARQL__CONTACT_EMAIL | bernt.popp@charite.de | Contact in the User-Agent (UniProt etiquette) |
UNIPROT_LINK_SPARQL__TIMEOUT | 30 | Per-request timeout (s) |
UNIPROT_LINK_SPARQL__DEFAULT_LIMIT | 50 | Auto-LIMIT for unbounded SELECTs |
UNIPROT_LINK_TRANSPORT | unified | unified / http |
UNIPROT_LINK_PORT | 8000 | Server port |
Development
See AGENTS.md for conventions (notably the SPARQL / QLever discipline that
keeps queries off the timeout cliff) and docs/ for architecture and usage.
research/verify_queries.py validates every query builder against the live
endpoint.
Disclaimer
Research use only; not for clinical decision support, diagnosis, treatment, or patient management. UniProt data is licensed CC BY 4.0.
License
MIT