Connect your AI

Give the AI you already trust a seat at your research bench. Once connected, your assistant can read the records you show it, capture what records say as AI-origin claims awaiting your review, keep your research log as it works, and suggest where to look next. Read-only calls and ordinary organizational writes can happen directly; evidence claims stay pending and gated changes stay proposals. Heartwood runs a local MCP server — stdio, and Streamable HTTP on 127.0.0.1 — so your evidence goes only to the model you personally chose, never through us.

Claude Desktop is the one client Heartwood can configure for you with a single click, from inside the app (Settings → Connect) — that's the only one-click path, and the only client we've verified end to end. Every other client below is a short manual setup using that client's own standard MCP configuration mechanism; we've written each one up from that vendor's own docs, but haven't tested it ourselves, so treat the vendor's docs link as the final word if anything looks off.

{
  "mcpServers": {
    "heartwood": {
      "command": "heartwood",
      "args": ["mcp"]
    }
  }
}

claude mcp add heartwood -- heartwood mcp

# or, to make it available in every project:
claude mcp add heartwood --scope user -- heartwood mcp

codex mcp add heartwood -- heartwood mcp

[mcp_servers.heartwood]
command = "heartwood"
args = ["mcp"]

gemini mcp add heartwood heartwood mcp

{
  "mcpServers": {
    "heartwood": {
      "command": "heartwood",
      "args": ["mcp"]
    }
  }
}

Why tools/list shows three tools

Heartwood's tools/list exposes exactly three bootstrap tools: search, describe, and invoke. Search for the task you need, describe the selected operation's exact schema, then call that operation through invoke. The generated MCP safety matrix lists every operation and its effect.

Three things to try first

Talk to your assistant the way you'd brief a research helper — plain language, real questions:

• “Here's a photo of my great-grandmother's marriage record — read it and capture what it says so I can review the claims.”
• “What do we actually know about Arthur Fernwood so far, and which parts are well-sourced?”
• “I'm stuck on who Eliza's parents were. Look at what we have and suggest where to search next.”

Evidence claims captured by your assistant land in a visibly separate pending lane, next to the source they came from, waiting for your review. Gated corrections and other proposal-only calls likewise wait for a person; the assistant cannot approve either lane itself. Read-only work and the direct-write tools identified in the MCP safety matrix do not create fake approval work. When you want more ideas, the prompt cookbook has a full set of starter prompts for real research sessions.

Optional: give your assistant a research discipline

Once connected, your assistant already knows how to call every Heartwood tool — that part needs nothing extra. If you want it to also follow a genealogist's research discipline (state a question, build a source plan, log every search, be explicit about what's still unconfirmed) rather than improvising, see the Heartwood Genealogical Proof Standard (GPS) skill: a short, plain-text guide that tells an AI assistant how to research the way a professional genealogist would, with wiring instructions for the assistant you use. It's entirely optional — Heartwood's connection works exactly the same without it — and you can edit or replace it with your own approach at any time.

Let your assistant read the docs itself

Your assistant can learn everything Heartwood offers without you explaining it. Give it this address: heartwood.family/llms.txt — a machine-readable index of these docs, built for AI assistants to navigate on their own. For the full catalogue of what it can call, point it at (or browse yourself) the MCP reference.