GeneBe

Why

Language models are bad at remembering the clinical status of a specific variant. They sound confident and are often wrong. GeneBe MCP lets the assistant look the variant up and answer from data: a transparent, points-based ACMG classification with every criterion explained, together with population frequencies, predictor scores, ClinVar and transcript consequences.

"Is NM_005577.4:c.4195A>C benign?"
"What is chr6-160585140-T-G? Explain why it is classified that way."
"Convert rs41272110 to hg38 and list its consequences on each transcript."

For example, chr6-160585140-T-G (rs41272110, LPA) comes back as Benign, −9 points with criteria BA1 and BP4, plus the data behind them and a link to the GeneBe page.

What you get

• Any notation in: genomic coordinates, HGVS, rsID or SPDI; hg19, hg38 and T2T supported through liftover.
• Verdict you can audit: ACMG classification with points and the reasoning for each criterion, not a black-box label.
• Evidence in one call: gnomAD frequencies, in-silico predictors, ClinVar, per-transcript consequences (HGVS c./p., impact), gene details with MANE Select and ClinGen gene–disease validity.
• Compact answers: large lists are capped, so one call costs roughly 1–3k tokens, and every result also carries structured data that follows the tool's output schema.
• Links back to GeneBe, so the human can check the source.

Connect in a minute

No sign-up, no API key. The server is remote; you only add its URL.

claude mcp add --transport http genebe https://mcp.genebe.net/v1/mcp

{ "mcpServers": { "genebe": { "url": "https://mcp.genebe.net/v1/mcp" } } }

{ "servers": { "genebe": { "type": "http", "url": "https://mcp.genebe.net/v1/mcp" } } }

curl -s -X POST https://mcp.genebe.net/v1/mcp \  -H 'Content-Type: application/json' -H 'Accept: application/json, text/event-stream' \  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"annotate_variant","arguments":{"variant":"chr6-160585140-T-G"}}}'

Tools

The default genome build is hg38.

Trust and privacy

• Read-only. No tool changes anything, anywhere.
• No account, no key needed to use the hosted server.
• Single variants only. No patient data, no VCF files, no samples are sent or stored.
• Nothing is stored and variant input is not logged by the MCP server.
• Transparent classification. Results come from GeneBe's points-based ACMG implementation; the reasoning is returned with the verdict, so you and the assistant can check it.
• Open source under the MIT license; you can read exactly what the server does, or run your own copy.
• Research use only. This is not a medical device and not a diagnosis. Classifications must be reviewed by a qualified specialist before any clinical use.

Roadmap

Today GeneBe MCP answers questions about single variants. We plan to grow it into a workspace your assistant can use on your own data. None of the items below is available yet, and the order may change.

• Sign-in, so the assistant can act on your GeneBe account
• Your own samples: view and edit sample details and annotations
• Search your cohort: find a variant across your own set of samples
• Upload samples from the assistant
• Supervise computations: start analyses, follow their progress and read the results

Ideas or requests? Open an issue or write to us via genebe.net.

Data sources

Results combine data from GeneBe, gnomAD, ClinVar, UniProt, ClinGen and dbNSFP predictors. Please check each source's license terms before redistributing the data.

Contributing and self-hosting

Want to run your own instance or change the code? See DEVELOPMENT.md. Issues and pull requests are welcome.