Przejdź do głównej zawartości

GeneBe MCP now reads your samples — sign in and ask about your own variants

Piotr
GeneBe maintainer

TL;DR: GeneBe MCP can now work with your own data. Connect your assistant to the new signed-in URL https://mcp.genebe.net/v1/signed-in/mcp (the public one without sign-in stays), sign in to GeneBe in your browser, and ask about your samples: find them, read their phenotype, and query the variants inside, for example "rare variants in LMNA in sample X". Everything is read-only, and you decide which permissions the assistant gets.

A week ago I introduced GeneBe MCP, which lets an AI assistant look up a single variant and answer from data instead of guessing. At the end I wrote that the next step was working with your own samples. That step is done.

What the assistant can do now​

After signing in, the assistant works under your GeneBe account and can:

  • find your samples (your own and those shared with your organizations): by words, project, sequencing id, experiment type (WGS, TRIO, ...), status, labels or creation date,
  • show a sample: status, labels, trio or pair members, input files and quality metrics,
  • read the phenotype: free text, HPO terms, diagnosis and sex, so it can relate the clinical picture to the variants,
  • list the result files of a sample and describe the columns of a variant table,
  • query the variants in a table with a SQL-like filter, for example gene = 'LMNA' AND gt <> 'REF/REF' AND AF_gnomad4 < 0.01.

Try asking:

  • "List the pathogenic, rare variants in the GJB2 gene in my sample." This is the question that matters most: the assistant picks the sample, checks what the columns mean, filters by gene, ACMG classification and population frequency, and lists what is left.
  • "Which of my trios created in the last 30 days are ready to view?"
  • "In sample smp_... list rare, non-reference variants in LMNA and tell me which look pathogenic."
  • "Does the phenotype of this sample fit any of the variants you found?"

The assistant first reads what the columns mean, then builds the filter, and a refused filter comes back with an explanation so it can fix it and retry. You can combine it with the public tools: take a variant from your sample and ask for its full ACMG reasoning.

You stay in control​

You sign in on genebe.net, never in the assistant, and the assistant never sees your password. The consent screen lists four separate permissions and you can untick any of them:

  1. public variant and gene lookups,
  2. the list of your samples and their basic details,
  3. clinical data (phenotype, diagnosis, sex),
  4. the variant tables of your samples.

Every tool is read-only: the assistant cannot change, delete or upload anything. You can withdraw access at any moment in Settings → Connected Apps.

Please keep in mind​

  • Sample names and sequencing ids are sent to the assistant and can identify a person. What the assistant reads becomes part of your conversation with the AI service you use, so check that this is allowed for your data.
  • Clinical data and variant rows are patient data. Grant the permissions you need and nothing more.
  • GeneBe is for research use only; classifications need review by a qualified specialist.

The MCP server itself stores nothing about your questions and does not log variant input.

How to connect​

There are two URLs, and you can add both:

  • public, no account: variant and gene descriptions only, no access to your samples or private data,
  • signed-in, with your GeneBe account: the same, plus your own samples.

In Claude Code:

# public: variant and gene descriptions, no account, no access to your samples or private data
claude mcp add --transport http genebe-public https://mcp.genebe.net/v1/mcp

# signed-in: the same, plus your own samples; you sign in with your GeneBe account
claude mcp add --transport http genebe-signed-in https://mcp.genebe.net/v1/signed-in/mcp

For the signed-in one, run /mcp, choose genebe-signed-in and Authenticate. Claude, ChatGPT, Cursor and VS Code work too; step-by-step instructions are in the documentation. Your samples are patient data, which is why reading them needs a sign-in and permissions you approve, while the public URL stays open to everyone.

What is next​

Still on the list: editing sample details, searching for a variant across all your samples at once, uploading samples, and starting and supervising analyses. The order may change, so tell me what would help you most.

The code is open source: github.com/genebe-net/genebe-mcp 🙏