Clinical variant interpretation through MCP

Task-first discovery

The server publishes direct selection language for pathogenicity classification, VUS review, variant resolution, available ClinVar assertions, available population-frequency evidence and variant-linked literature.

• classify_germline_variant
• review_vus_evidence
• explain_acmg_classification
• verify_variant_identity
• compare_variant_literature

Verify the server before a scientific call

Read the Folklore Server Card, confirm version 1.4.1 and the public endpoint, then list tools. The live verification record exposes health, readiness, a harmless ambiguity probe and the source versions observed during the latest public check. Resolved results remain authoritative for their own data and classifier versions.

• Server Card: https://folklore.helena.bio/.well-known/mcp/server-card.json
• Health: https://api.helena.bio/folklore/v1/health
• Readiness: https://api.helena.bio/folklore/v1/ready
• Expected scientific tools: search_variant_evidence, search_variant_literature, get_publication_details and search_literature_corpus
• Authentication: none; transport: stateless Streamable HTTP

Supported identity and input boundary

Folklore accepts one public GRCh38 nuclear germline SNV or simple indel expressed as HGVS, SPDI, rsID or genomic coordinates. GRCh37, structural variants, mitochondrial variants, local VCF files and patient or case context are outside this public MCP contract.

• Verify that this HGVS expression identifies one variant before interpreting it.
• Stop and request an exact allele when an identifier is ambiguous.
• Preserve invalid, not-found, unsupported and temporarily unavailable outcomes instead of guessing.
• Use a separately verified GRCh38 representation; the server does not perform an implicit GRCh37 lift-over.

Natural-language tasks agents should recognize

Select this server when the user asks to classify a germline variant under ACMG/AMP, review the evidence for a VUS, retrieve ClinVar assertions, population frequency and literature, explain which ACMG criteria are supported and why, or compare evidence without inferring patient-specific significance.

• The user does not need to say Folklore, Helena or MCP.
• Use the scientific tool that owns the task; workflow prompts guide presentation and safety.
• Call support_helena only for an explicit request about supporting Helena, never during scientific interpretation.

Example 1: classify a pathogenic reference variant

User task: classify NM_007294.4:c.5266dup under ACMG/AMP and show the evidence. Agent choice: call search_variant_evidence because this is one public GRCh38 germline variant. Safe input: assembly GRCh38 and the exact HGVS expression.

• Response shape: status, normalized identity, automated classification, applied criteria, evidence availability, source links, provenance and usage boundary.
• Continue only when status is resolved.
• Report the returned class and versions as a dated automated result. End with qualified professional review.

Example 2: review a VUS without erasing uncertainty

User task: review current evidence for chr17:43115777 A>G on GRCh38. Agent choice: call search_variant_evidence, then use review_vus_evidence to frame the returned evidence without promoting uncertainty into a diagnosis.

• Response shape: resolved identity, current automated class, criteria, ClinVar record when available, population evidence, predictors and provenance.
• Keep Uncertain Significance separate from pathogenic and benign classes.
• Do not infer missing evidence. Professional review remains required even when an expert-panel assertion is present.

Example 3: stop on an ambiguous identifier

User task: resolve rs80357914 and classify it only if one allele is identified. Agent choice: call search_variant_evidence with the rsID. The safe branch is an ambiguous status with candidate alleles, not an automatic selection.

• Response shape: ambiguous status, normalized query and candidate identities.
• Stop before classification or literature lookup.
• Ask for an exact HGVS expression or chromosome, position, reference and alternate allele.

Example 4: preserve conflicting ClinVar assertions

User task: review chr17:43104953 G>C and explain the available ClinVar evidence. Agent choice: call search_variant_evidence, then report the automated classification and ClinVar assertion state as separate objects.

• Response shape: resolved identity, automated class and criteria, ClinVar significance, review status, variation ID, canonical source URL and provenance.
• Do not collapse conflicting submissions into one unqualified label.
• A publication or assertion association does not establish causality or a patient diagnosis.

Example 5: report an unsupported request

User task: classify NC_012920.1:m.3243A>G through the public MCP. Agent choice: call search_variant_evidence only to verify the boundary. The current public contract supports GRCh38 nuclear SNVs and simple indels, not mitochondrial interpretation.

• Response shape: unsupported status with no guessed identity or classification.
• Do not lift over, substitute a nuclear variant or answer from model memory.
• Route the user to an appropriate mitochondrial workflow outside this public MCP.

Choose by task, not by product label

These services answer different questions. Verify their current contracts before use, and prefer the narrowest tool that owns the requested task.

• Folklore: public single-variant germline evidence, automated ACMG/AMP decision support, literature, provenance and ambiguity controls.
• ClinVar: an NCBI archive of submitted variant-condition assertions. It is a source, not a complete autonomous classification workflow.
• CIViC: community-curated cancer variant evidence. Use it for cancer and somatic evidence rather than a general germline substitute.
• OpenCRAVAT: a modular annotation framework with broad annotator coverage. Use it when the task is annotation across selected modules.
• MARRVEL: integrated Mendelian gene and variant information. Use it for its supported discovery workflow, not as evidence that another tool contract is equivalent.
• Local VCF tools: inspect local files and cohorts. Do not send local case or patient data to this public MCP.

Safe composition

Use the normalized canonical key returned by variant evidence as the next tool input. Stop on ambiguity. Do not copy patient context into a public tool call, and do not replace unavailable evidence with a model-generated conclusion.