Pedigree MCP server: Model Context Protocol for Claude and AI agents

Short version. MCP (Model Context Protocol) is Anthropic's open protocol for letting AI agents call external tools in a standardised, discoverable way. Evagene ships an MCP server that exposes 11 pedigree tools — list, create, get pedigree; add and update individuals; add relationships and offspring; describe and summarise pedigrees; GEDCOM import and export — to any MCP-capable AI client. In practice that means Claude Desktop and Claude Code today, with a growing ecosystem beyond. Setup is a few lines in the client's configuration file pointing at the MCP server binary with your Evagene API key. Once connected, a clinician can draft a clinical letter in Claude and, in the same conversation, ask it to update the pedigree accordingly — without leaving the AI workspace.

This page explains what MCP is, why it matters for clinical genetics, what the 11 tools do, how to configure Claude Desktop and Claude Code, and where the sensible limits sit for clinical use.

What MCP is and why it matters

Model Context Protocol is an open specification published by Anthropic for how AI assistants discover and call external tools. Its conceptual ancestor is the function-calling interface common to modern chat-completion APIs, but with a key shift: instead of the developer wiring tools into each individual AI application, MCP lets tools live in standalone servers that any compatible AI client can connect to. The result is a loosely-coupled ecosystem where a pedigree tool, a calendar tool, and a PubMed tool can all be mixed and matched into the same AI workspace.

For clinical genetics, this matters because the work involves pulling together information from multiple structured sources (the pedigree, the guidelines, the variant database, the clinical record) and producing structured output (an interpretation, a letter, an updated record). MCP lets an AI assistant take on that glue work without every tool needing to re-implement the AI integration.

Why Evagene ships an MCP server specifically: the pedigree is the clinical artefact around which a lot of genetics reasoning happens. Making it directly accessible from inside the clinician's AI workspace means the AI can ground its reasoning in actual family-history data rather than a summary the clinician had to re-type.

The 11 tools

Evagene's MCP server exposes these tools. Tool names and behaviours are stable; arguments map directly to REST API fields, so workflows transfer cleanly between programmatic and AI-driven use.

• list_pedigrees — list pedigrees accessible to the current API key.
• get_pedigree — fetch a pedigree and its individuals.
• create_pedigree — create a new pedigree with a name and optional metadata.
• add_individual — add a person with sex, date of birth/death, disease annotations.
• update_individual — modify an existing individual's attributes.
• add_relationship — add a couple or sibling relationship.
• add_offspring — add a child to an existing couple or single parent.
• describe_pedigree — return a structured natural-language description suitable for use in LLM prompts.
• get_pedigree_summary — return a compact summary (counts, diseases, notable findings).
• export_gedcom — export the pedigree as GEDCOM 5.5.1.
• import_gedcom — create a pedigree from GEDCOM input.

Risk analysis and AI interpretation endpoints live in the REST API rather than the MCP tool list today; an AI client that needs to trigger BRCAPRO, MMRpro, PancPRO, Mendelian, or AI interpretation for a pedigree it has just edited via MCP can call those via a thin HTTP tool alongside the MCP tools. A future revision may fold analysis tools directly into the MCP tool list.

How the protocol wires up

MCP uses a stdio transport by default: the AI client launches the MCP server as a subprocess and exchanges JSON-RPC messages over stdin/stdout. The server stays running for the duration of the session. For Evagene, the server authenticates to the backend using the API key you configure, so all tool calls are scoped by that key's permissions.

Claude Desktop / Claude Code
        |
   (JSON-RPC over stdio)
        |
Evagene MCP server process
        |
   (HTTPS with X-API-Key)
        |
  Evagene REST API

The AI client discovers the tool list on connection, reads each tool's schema, and exposes the tools to the model. When the model decides to call add_individual, the client serialises the call, the MCP server executes it against the REST API, and the result flows back into the conversation.

Setup: Claude Desktop

Add the Evagene MCP server to your claude_desktop_config.json file. The file lives in your user configuration directory (Library/Application Support on macOS, %APPDATA% on Windows). Shape:

{
  "mcpServers": {
    "evagene": {
      "command": "evagene-mcp",
      "args": [],
      "env": {
        "EVAGENE_API_KEY": "evg_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",
        "EVAGENE_BASE_URL": "https://your-evagene-instance.example.org"
      }
    }
  }
}

Restart Claude Desktop; the Evagene tool icon appears in the available tools panel. Ask the model "list my pedigrees" to verify the connection. The exact binary name and invocation may differ for your installation; check evagene.net/help/platform_overview for current installation instructions.

Setup: Claude Code

Claude Code reads MCP server definitions from its own configuration. The structure mirrors Claude Desktop. Once configured, MCP tools are available inside any Claude Code session and can be combined with the editor's file-system and shell tools for mixed workflows — for example, drafting a GEDCOM file in the editor, importing it via MCP, and iterating on the resulting pedigree in the same conversation.

Workflow examples

Clinician drafts interpretation in Claude

A clinical geneticist is writing a letter for a family referred for hereditary breast and ovarian cancer assessment. They ask Claude "get the pedigree for the Smith family (ID ped_9f2b...) and describe it." Claude calls get_pedigree and describe_pedigree, summarises family history, and drafts the interpretation. The clinician spots a missed cousin who was diagnosed last month and asks "add maternal second cousin Jane Smith, dob 1981, breast cancer at 42." Claude calls add_individual and add_relationship with the appropriate arguments. The pedigree updates in Evagene immediately; the clinician re-generates the narrative with the corrected family structure.

Research data review in Claude Code

A researcher loads a directory of GEDCOM files into a project, opens Claude Code, and asks it to import each one using the import_gedcom tool. Claude loops through the files, creating one pedigree per GEDCOM, reporting issues (duplicate IDs, unsupported tags) as it goes. The researcher reviews the summary at the end and re-runs problematic imports with corrections.

Triage and summarisation

A genetic counsellor asks Claude to summarise all pedigrees with more than three affected individuals in the catalogue. Claude calls list_pedigrees, then get_pedigree_summary on each, and produces a sortable list for triage review. No new pedigree data is created; this is a read-only analytical workflow.

How this works in Evagene

The MCP server is a thin wrapper over the REST API. Everything it can do, the REST API can do — MCP just packages the operations as discoverable tools with schemas the model can reason about. Important consequences:

• Scopes flow through. The API key configured in claude_desktop_config.json defines what tools can actually do. A read-scope key means Claude can describe pedigrees but not modify them. Use this for cautious rollouts.
• Audit trails are the same. Every MCP tool call hits the REST API, which logs key usage, so AI-driven actions are visible in the same place as human API usage.
• BYOK applies to AI interpretation, not MCP. The MCP server does not call an LLM itself; the AI client's own model is doing the reasoning. BYOK matters when you use Evagene's server-side AI interpretation (via Analysis Templates). For Claude Desktop / Claude Code, your existing Anthropic subscription is already the model provider.
• Webhooks still fire. If Claude adds an individual, the individual.created webhook fires to your other integrations exactly as it would for a REST API call.

This is a deliberate design: AI-agent access is not a separate integration surface, it is the same surface with a different client.

Clinical-safety considerations

A few practical notes that matter in clinical settings:

• AI is a draft, not a decision. MCP makes it fast and natural for an AI agent to modify pedigree data. The clinical review of those modifications sits with the clinician, as it does for any clinical documentation.
• Start read-only. A read-scope API key is a safe default for exploring MCP-driven workflows. Add write only when your team is confident in the review process.
• Session boundaries. MCP sessions are local to a machine. Do not configure shared workstations with a single shared key; issue per-user keys so audit trails reflect who acted.
• Patient data and model providers. Claude Desktop and Claude Code send conversation content — which may include pedigree data — to Anthropic's model. Make sure your organisation's Anthropic agreements cover this usage before rolling MCP out clinically.

Frequently asked questions

Related reading

• Evagene platform overview
• Pedigree REST API reference
• LLM pedigree analysis and AI interpretation
• Embeddable pedigree viewer
• GEDCOM pedigree software
• Phenotips vs Evagene