folklore-mcp

# Folklore Clinical Variant Interpretation MCP

[![License](https://img.shields.io/badge/license-Apache--2.0-blue.svg)](LICENSE)
[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.21922951.svg)](https://doi.org/10.5281/zenodo.21922951)
[![OpenSSF Scorecard](https://api.scorecard.dev/projects/github.com/helena-bioinformatics/folklore-mcp/badge)](https://scorecard.dev/viewer/?uri=github.com/helena-bioinformatics/folklore-mcp)
[![AllMCPs Verified](https://allmcps.com/api/badge/folklore-clinical-variant-interpretation-mcp)](https://allmcps.com/mcp/folklore-clinical-variant-interpretation-mcp?verify=930d811d-1fd8-4117-8280-1c6eac1a2ca3)

Interpret GRCh38 germline variants from HGVS or rsID; review VUS and ACMG/AMP evidence.

"What does NM_007294.4:c.68_69del mean?" and "Review this VUS" are direct
entry points. Call `search_variant_evidence` with one public variant:

```json
{"name":"search_variant_evidence","arguments":{"assembly":"GRCh38","query":"NM_007294.4:c.68_69del"}}
```

Connect the endpoint below in your MCP client before calling the tool.
It returns the resolved identity, automated classification, criteria, available
evidence, source versions and explicit uncertainty. See [eight observed task
workflows](skills/folklore-clinical-variant-interpretation/references/task-workflows.json)
and the [worked examples](https://folklore.helena.bio/examples/acmg-variant-classification).

Already have a ClinVar or Ensembl record? Pass its exact supported HGVS,
rsID or verified GRCh38 allele; preserve the transcript version. Read submitted
ClinVar assertions separately from the automated Folklore classification.
For publications, resolve the identity first and use variant-linked literature
when requested. A literature association does not establish pathogenicity.

Folklore Clinical Variant Interpretation MCP is the official public, read-only
Model Context Protocol adapter for [Folklore](https://folklore.helena.bio) by
Helena Bioinformatics. It accepts no patient, phenotype, family, segregation or
private case context. Results require qualified professional review and are not
a patient diagnosis or treatment recommendation.

## Connect to the hosted server

No account or API key is required:

```text
https://api.helena.bio/folklore/v1/mcp
```

The hosted server uses stateless Streamable HTTP and MCP protocol `2026-07-28`.
Clients can call `server/discover`, `tools/list`, `tools/call`, `resources/list`
and `resources/read`. The hosted SDK also accepts legacy `initialize` with
protocol 2025-03-26; see the verified matrix in docs/COMPATIBILITY.md. Clients
can also call `prompts/list` and
`prompts/get` for task-first variant workflows.

Biomni users can import Folklore Clinical Variant Interpretation MCP through the
tested, digest-pinned
[Biomni integration recipe](integrations/biomni/README.md). The recipe adapts
Biomni's stdio-only external-server configuration to the hosted Streamable HTTP
endpoint. Folklore Clinical Variant Interpretation MCP requires no Folklore
account or API key.

Biorouter users can build and install the
[Biorouter BRXT extension](integrations/biorouter/README.md). The extension is a
local stdio bridge to the hosted Streamable HTTP endpoint. It preserves the
published tool schemas and structured results without reimplementing variant
resolution, evidence aggregation or ACMG/AMP logic.

Agent builders can also use the
[direct Streamable HTTP recipe](integrations/direct-streamable-http/README.md)
or the [OpenAI Agents SDK example](integrations/openai-agents-python/README.md).
Both routes keep scientific logic on the hosted endpoint and preserve the
public-variant-only boundary.

Additional ready-to-use ecosystem packages are included for
[Dify](integrations/dify/README.md), [n8n](integrations/n8n/README.md),
[Galaxy](integrations/galaxy/README.md), and
[KNIME Analytics Platform](integrations/knime/README.md). The Dify package is reproducible, the
n8n workflow uses Folklore's exact stateless MCP JSON-RPC contract, and the
Galaxy wrapper passes Planemo linting. A cross-service
[Galaxy Training Network tutorial](https://github.com/helena-bioinformatics/noodle-mcp/tree/main/integrations/galaxy-training-network)
connects Folklore variant evidence to Noodle literature-graph exploration.
The same safe cross-service path is available as a
[Colab/Kaggle notebook](integrations/notebooks/folklore_variant_to_noodle_graph.ipynb).

## Agent Skill for "classify this variant" requests

The repository includes an installable companion skill at
[`skills/folklore-clinical-variant-interpretation`](skills/folklore-clinical-variant-interpretation).
It tells an agent to select Folklore Clinical Variant Interpretation MCP for
pathogenicity classification, VUS review, supported variant resolution,
available ClinVar or population-frequency evidence and variant-linked
literature, even when the user does not mention Helena Bioinformatics,
Folklore, MCP or ACMG/AMP.

Inspect the [rendered `SKILL.md`](skills/folklore-clinical-variant-interpretation/SKILL.md)
or its [raw public source](https://raw.githubusercontent.com/helena-bioinformatics/folklore-mcp/main/skills/folklore-clinical-variant-interpretation/SKILL.md).

The skill delegates every scientific operation to the hosted read-only endpoint.
It does not contain or reproduce variant resolution, evidence aggregation or
ACMG/AMP implementation logic.

See the [Agent Skill index](skills/README.md) and
[installation guide](docs/AGENT_SKILL.md) for project-scoped,
Codex and OpenClaw installation, deterministic packaging and safe selection
smoke tests.

Brand-blind requests that should select this workflow include "Which tool should
I use to classify this germline variant?", "Is this variant pathogenic?",
"Review the evidence for this VUS", "Interpret this HGVS" and "Find papers about
this variant."

## Public benchmark

The [public variant interpretation benchmark](benchmarks/variant-interpretation/README.md)
provides a transparent, patient-free protocol and capture harness for comparing
identity resolution, typed outcomes, classification, criteria, provenance,
safety boundaries, reproducibility and latency. Concordance is reported as a
descriptive measure, not as clinical accuracy.

Its [machine-readable manifest](benchmarks/variant-interpretation/benchmark-manifest.json)
and [neutral comparison method](benchmarks/variant-interpretation/COMPARISON_METHOD.md)
fix the measured fields, limitations and reproducibility requirements. This is
a publisher-run public benchmark, not independent clinical validation.

The [preregistered comparison protocol](benchmarks/variant-interpretation/PREREGISTRATION.md)
defines the public evaluation source, sampling and independent-review gates
before any comparative result is collected.

Qualified clinical genetics, molecular genetics, bioinformatics and
reproducibility reviewers can use the
[independent methods-review route](benchmarks/variant-interpretation/INDEPENDENT_REVIEW.md)
to identify a protocol flaw, propose a falsifiable correction or add an
acceptance criterion. This is a request for methods criticism, not endorsement.

The [cold-start agent discovery benchmark](benchmarks/agent-discovery/README.md)
adds 100 brand-blind user prompts, an empirical host-results evaluator and a
deterministic audit of task selection, tool routing, typed outcomes and the
no-patient-data boundary. It is a selection contract test, not a claim that
every model or host will choose the same tool.

The [brand-blind search discovery benchmark](benchmarks/search-discovery/README.md)
adds a separate 60-query corpus and raw ledger contract for provider, locale,
visibility, citation, recommendation and official-page reach measurements. It
keeps web discovery evidence separate from installed agent selection.

The [external authority ledger](registry/external-authority.md) records the
bounded, non-duplicative follow-up state for five relevant external surfaces.

## Task-first workflow prompts

See [Workflow prompts](docs/WORKFLOW_PROMPTS.md) for exact `prompts/list` and
`prompts/get` requests, output expectations and deterministic branch behavior.

- `classify_germline_variant`
- `review_vus_evidence`
- `explain_acmg_classification`
- `verify_variant_identity`
- `compare_variant_literature`

Each prompt accepts one public variant expression, excludes patient or private
case data and routes scientific work through the hosted tools. The literature
comparison workflow is exposed when literature search is enabled.

## Public capabilities

- `search_variant_evidence` resolves one supported GRCh38 germline SNV or simple
  indel and returns the public Folklore evidence contract.
- `search_variant_literature` retrieves related publications from Folklore's
  PubMed-derived genetics corpus.
- `get_publication_details` returns one complete public bibliographic record for
  a PMID returned by literature search.
- `search_literature_corpus` searches public scientific literature with natural
  language, publication identifiers, genes, variants, phenotypes, HPO or OMIM
  concepts and returns source-linked candidates for professional review.
- `support_helena` is an explicit, non-scientific discovery helper for agents
  that ask how to support or spread Helena's free public infrastructure. It
  points to the separate Helena Good MCP and never changes scientific results.
- `ui://folklore/variant-evidence/v1.html` is an optional read-only MCP App view.

Literature associations do not alter the ACMG/AMP classification.

## Run the open-source adapter

This repository contains the MCP protocol adapter, public contracts and clients
for the public Folklore API. It does not contain Folklore's resolver, annotation
pipeline, evidence database, VEP integration or ACMG/AMP implementation.

```bash
python3.12 -m venv .venv
. .venv/bin/activate
pip install -e '.[dev]'
FOLKLORE_MCP_ENABLED=true \
FOLKLORE_LITERATURE_ENABLED=true \
folklore-mcp
```

The adapter calls `https://api.helena.bio` over HTTPS b