Loading ToolBench
Look up genes, fetch sequences, predict variant consequences, find orthologs, and retrieve cross-database xrefs from Ensembl REST via MCP. STDIO or Streamable HTTP.
Strong tool definitions with excellent descriptions (avg 250+ chars), complete input schemas with types and descriptions, and clear parameter constraints via enums. All 6 tools follow verb_noun naming (ensembl_get_*, ensembl_list_*, ensembl_predict_*). Output schemas are documented in descriptions. Minor gaps: no explicit error handling guidance, no per-tool security annotations despite READ_ONLY risk labels, and output field documentation could be more structured. Descriptions are LLM-optimized and include use-case context ('what is the mouse equivalent of human TP53?'). Parameters consistently have types, descriptions, and constraints (enums, min/max, defaults). Batch operations supported (lookup_gene accepts ids/symbols arrays up to 20 items). Tool composition is sound, tools chain naturally (lookup_gene → get_sequence/get_xrefs/get_homology).
Find orthologs and/or paralogs of a gene across species. Returns each homolog's stable ID, species, homology type (ortholog_one2one, ortholog_one2many, paralog_many2many, etc.), perc_id (percent identity), perc_pos (percent positives), and taxonomy level. Essential for cross-species research — for example, "what is the mouse equivalent of human TP53?" or "how conserved is BRCA2 across mammals?". Provide either symbol + species or a stable gene ID. Target species can be filtered to a single species or left open to return all available homologs.
Fetch the DNA, cDNA, CDS, or protein sequence for a gene, transcript, protein, or genomic region. Returns the sequence with its stable ID, molecule type, and character count — large sequences are returned in full but the length is stated so callers can budget context. The type parameter selects which sequence is fetched: genomic (default, includes introns), cdna (spliced transcript), cds (coding sequence only), protein. For region mode, set id to a region — either species:chr:start-end (e.g. homo_sapiens:13:32315086-32400268) or a bare chr:start-end with species set (e.g. id 13:32315086-32400268, species homo_sapiens). Protein sequences require a transcript or protein stable ID (ENST…/ENSP…), not a gene ID — use ensembl_lookup_gene with expand_transcripts=true to get the canonical transcript ID first.
Retrieve cross-database references for a gene or feature — HGNC, UniProt, EntrezGene, OMIM, RefSeq, Reactome, and others. Returns each xref with its database name, primary ID, display ID, and description. The dbname filter narrows to specific databases; omit to return all xrefs. IDs returned here chain to protein (pubchem via UniProt), literature (pubmed via PubMed IDs), disease (OMIM via MIM_GENE), and pathway (Reactome) resources. Requires an Ensembl stable ID — use ensembl_lookup_gene to get the ENSG… ID first. Common dbname values: HGNC, Uniprot_gn, EntrezGene, MIM_GENE, RefSeq_mRNA, RefSeq_peptide, Reactome, GO (Gene Ontology), ChEMBL.
No explicit error handling guidance in tool descriptions. Tools do not document what happens on invalid input (e.g., invalid species, malformed variant notation, gene not found) or how LLMs should recover.
Output schemas documented in descriptions but not formally structured. No explicit field-level documentation (e.g., what does 'perc_id' mean in homology results? What are valid consequence terms in VEP output?).
Tool descriptions include example values (e.g., 'BRCA2', 'TP53', 'rs334') which LLMs may reuse literally in real calls. Should use parameter constraints (enums, patterns) instead.
Inferred effective spec: 2026-07-28+.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | A | 83 | 2026-07-28+ | v2 |
List species supported by Ensembl with display name, common name, assembly, taxon ID, and division. Required discovery step — species names like homo_sapiens are opaque to non-biologists and are the input format every other Ensembl tool expects. Filter by division to select one; use nameContains to find a species by partial name match. With no division, returns the endpoint default division — the vertebrates (~356 species on the default GRCh38 endpoint); pass a division to list that division.
Resolve a gene by symbol + species (or by stable ID) to its Ensembl ID, genomic location (chr:start-end:strand), biotype, description, and transcript list. Entry point for most workflows — the stable ID and coordinates returned here are inputs to other tools. Accepts both symbol lookup (BRCA2 + homo_sapiens) and direct ID lookup (ENSG00000139618). Supports batch lookup of up to 20 IDs or symbols in one call via the ids or symbols field. Provide exactly one of symbol, id, ids, or symbols. For symbol lookups species defaults to homo_sapiens (override for other organisms); for ID lookups species is not needed. Use ensembl_list_species to discover valid species names.
Predict the functional consequences of a sequence variant using the Ensembl Variant Effect Predictor (VEP). Accepts three input formats: HGVS notation (transcript-relative, e.g. ENST00000380152.8:c.2T>A, or genomic, e.g. 13:g.32316462T>A); region+allele (chr:start:end:strand/allele, e.g. 1:65568:65568:1/T); and a dbSNP rsID (e.g. rs334). Returns the most severe consequence term, affected transcripts and genes, impact level (HIGH/MODERATE/LOW/MODIFIER), and any colocated known variants with clinical significance. HGVS input: provide the full notation including transcript version for best results. Region+allele input: Ensembl normalizes chromosome names and canonical vertebrate output omits the chr prefix (a chr-prefixed name is also accepted). By default the response caps transcript consequences (max_transcript_consequences) and per-variant PubMed IDs (max_pubmed_ids_per_variant) to keep large VEP results compact — well-studied variants like rs334 otherwise carry 60+ consequences and 100+ citations. Truthful totals are always reported; set a cap to 0 (or include_all_colocated_pubmed=true) to retrieve the full set.
No confirmation or dry-run pattern for tools that may return large result sets (e.g., ensembl_get_homology with max_results=0 can return 150+ items). No guidance on context window impact.
Tool annotations present (READ_ONLY risk labels) but not formalized as idempotentHint/readOnlyHint in schema. Limits agent reasoning about retry safety.