Skip to content

Privacy Policy

What this software is

AlphaFold Sovereign MCP is local-first, on-premises software. It is not a SaaS. The software runs on your infrastructure, under your control.

This document describes what the shipped server does with the values it is given. It makes no statement about how any law or regulation applies to a particular deployment.

Data This Software Processes

The tools accept scientific identifiers and search text:

  • UniProt accessions (e.g., P12345), gene symbols, and Ensembl gene IDs.
  • Variants in HGVS notation.
  • Disease and phenotype identifiers: MONDO IDs, HPO IDs, and ICD-10 codes.
  • Free-text search terms, such as a disease name.
  • Species names, numeric limits, thresholds, and option flags.

No tool has a field for a person's name, date of birth, contact details, or record number, and no tool accepts a sequence or a file. Input validation checks identifier formats and numeric ranges. It does not look for personal information, so the software cannot tell whether a variant or a search term relates to an identifiable person. Whatever is supplied is handled as the sections below describe: it is sent to the upstream service that answers the query, it can appear in the server log, and the result is returned to the MCP client that made the call. Deciding which values may be supplied is left to whoever runs the server.

Outbound Network Calls

In online mode (default), the tools that query an upstream service make HTTPS requests through one shared client. That client contacts the hosts below and refuses any request or redirect that leaves the configured origin.

Host Service Used for Sent in the request
alphafold.ebi.ac.uk AlphaFold DB Predicted structures, pLDDT, PAE, AlphaMissense scores UniProt accession
rest.uniprot.org UniProtKB Identity check for an explicitly requested isoform UniProt accession
rest.ensembl.org Ensembl REST Variant consequences (VEP), gene lookup, orthologs HGVS variant, gene symbol, Ensembl gene ID
eutils.ncbi.nlm.nih.gov ClinVar through NCBI E-utilities Variant classifications HGVS variant or gene symbol as the search term
gnomad.broadinstitute.org gnomAD Allele frequencies, gene constraint Variant ID (chromosome, position, alleles) derived from the HGVS input; gene symbol
api.platform.opentargets.org Open Targets Platform Target-disease evidence, tractability UniProt accession or gene symbol, Ensembl gene ID, disease ID or name
ontology.jax.org HPO Phenotype terms, disease and gene annotations HPO term ID, NCBI gene ID
www.ebi.ac.uk MONDO through OLS4 Disease terms, hierarchy, cross-references MONDO ID, disease search text, ICD-10, OMIM or Orphanet code
www.ebi.ac.uk ChEMBL Drug targets, approved drugs, indications UniProt accession, ChEMBL ID, disease ID or name
www.ebi.ac.uk PDBe Residue-level literature annotations, only when requested UniProt accession and residue number
api.disgenet.com DisGeNET Gene-disease association scores Gene symbol

Every request carries a User-Agent header naming this software, its version, and its repository URL. If you set NCBI_API_KEY or DISGENET_API_KEY, that key is sent to the corresponding service as a request parameter. As with any HTTPS request, the receiving service also sees the network address the request comes from.

These services are operated by third parties. This project does not control how they handle the requests they receive.

Offline mode

With ALPHAFOLD_OFFLINE set to 1, true, or yes when the server starts, the shared client refuses every request before a connection is opened, except to hosts you list in ALPHAFOLD_ALLOW_HOSTS, and the framework's update check described below is turned off. There is no bundled data set and no cache of upstream responses. A tool that needs an upstream service then fails with an error or returns a result that reports the affected lookups as failed or not assessed; it does not fall back to stored data. The knowledge-graph tools read only the local database (see Local Storage) and keep working.

Update check by the MCP framework

In online mode one start-up request is made outside that client. When the server starts, FastMCP, the framework it is built on, asks pypi.org whether a newer FastMCP release exists. The request carries no tool input, and FastMCP keeps the answer for 12 hours in a cache file in its own data directory. Set FASTMCP_CHECK_FOR_UPDATES=off to turn it off.

In offline mode the server turns this check off itself before it starts, whatever FASTMCP_CHECK_FOR_UPDATES says, and the process starts without any name lookup or connection. Both paragraphs describe FastMCP 4.1.4 and 4.1.8, the versions this was observed with.

Logs

This software writes structured log lines to its standard error stream and to no file. A line can contain the identifier or search text of the call it describes (for example a UniProt accession, a gene symbol, an HGVS variant, or a disease search term), the name of the upstream service, the request path, and error text. The program that launches the server decides whether standard error is displayed, stored, or discarded.

Local Storage

This software keeps one local SQLite database, the knowledge graph, in the per-user application data directory or at the path given in ALPHAFOLD_KG_PATH. The server creates it with empty tables: a new database holds no records, only a schema-version marker. No registered tool and no start-up step writes records to it. Records exist only if you, or code you add, write them through the storage API, or if the database comes from an earlier release that seeded illustrative rows, which this release leaves untouched. The knowledge-graph tools read the database and return what they find to the MCP client that calls them.

The database is ordinary SQLite. This software does not encrypt it and places no access restriction on it beyond the file permissions of your system.

Telemetry

None by default. The software does not call home, report usage, or transmit any telemetry, and it contains no telemetry code or exporter of its own.

FastMCP (4.1.4 and 4.1.8 were examined) includes OpenTelemetry instrumentation built on the OpenTelemetry API package alone. No OpenTelemetry SDK or exporter is a dependency of this software, so that instrumentation exports nothing unless you install an SDK and configure an exporter yourself. If you do, spans go to the endpoint you configured, and their content is determined by FastMCP, not by this project.

Audit Log

There is no automatic audit log. The database has a tool_invocations table and the storage API has a function, KnowledgeGraph.log_tool_invocation, that writes to it, but the shipped server never calls that function. In normal operation no tool call is recorded and the table stays empty.

If code you add calls it, each record holds the tool name, the complete input parameters and the complete result as JSON text, a SHA-256 hash of each, the duration, any error text, session and request identifiers, and timestamps. The records are not limited to hashes.

There is no retention setting. The software contains no expiry, purge, or rotation code, so whatever is written to the database stays there until you delete it. Nothing sends the database to another system. The dataset-export tool returns the tables it is asked for to the MCP client, and it accepts tool_invocations as a table name.

Personal and Health Data

This software is not designed for data about identifiable people. It has no de-identification step, no access control of its own, and no encryption of the local database, and it holds no certification for regulated use (see STATUS.md).

This document makes no claim about whether GDPR, HIPAA, or any other law applies to a deployment, or about what such a law would require.

Contact

Privacy questions: open a GitHub Discussion or, for sensitive issues, follow the coordinated-disclosure process in SECURITY.md.

Last updated: 2026-11-11