xhuman← Back

MCP Connector Privacy

Last Updated: 2026-09-07

This document describes how the Xhuman MCP server ("the Connector") handles your data when you connect it to Claude.ai, ChatGPT, Claude Desktop, or any other MCP-compatible AI client. It supplements — not replaces — the Xhuman Privacy Policy, which governs everything we do with your data on Xhuman itself.

By connecting your AI client to https://xhuman.com/api/mcp and authorizing access, you agree to the practices described below.

What the Connector does

The MCP Connector exposes tools that an authorized AI client can call on your behalf. Some of them read your data; some of them change it.

Tools that read:

  • search_mind — keyword/hybrid search across your wiki
  • read_article — fetch one article by slug
  • ask_mind — retrieval-augmented Q&A over your wiki
  • timeline — your chronological life timeline
  • list_index — metadata for every wiki article
  • get_graph — your wikilink knowledge graph
  • search_memories — semantic search over your atomic memories
  • list_memories — chronological slice of your memories

Tools that write:

  • ingest_source — add pasted text, or a URL we fetch, as a new source
  • record_fact — record a fact as an atomic memory
  • update_article — replace the body of one wiki article
  • delete_article — permanently delete a wiki article
  • update_memory — replace the text of one memory
  • delete_memory — delete a memory, hiding it from search immediately
  • correct_claim — revoke a wrong claim and record the correction

The writing tools unlock with the scopes you grant when you authorize a client. The standard connect flow grants the mcp scope, which unlocks all of them, including the two that delete; the authorization screen lists this before you approve. A client granted only the read scope gets the reading tools, and the server refuses its write calls.

Deleting an article is permanent. Deleting a memory removes it from search and retrieval immediately.

The Connector cannot send email, post messages, or take any action outside your own data. The only request it makes to anywhere else is ingest_source fetching a URL you asked it to add.

What data the Connector accesses

When an AI client calls a tool, the Connector reads from:

  • Your wiki articles and their bodies
  • Your atomic memories (deduped facts extracted from chats and voice)
  • Your timeline events and life eras
  • Your wikilink graph

When a client calls a writing tool, the Connector adds to or changes that same data, and ingest_source adds to your saved sources.

The Connector does not access:

  • Your raw conversation transcripts (chats, voice calls)
  • Your existing saved sources (PDFs, web clips, ingest history) — it can add a new source, but it cannot read back what is already there
  • Your account settings or billing data
  • Anyone else's data

Audience binding

Tokens issued through OAuth are bound to your [self] audience by default — meaning the Connector sees your data exactly as you do. We support narrower audience bindings (e.g. [everyone], where the Connector only sees public claims), but the standard "Connect" flow issues a self-bound token. You can mint a custom-audience token from your MCP settings.

What we send to third parties

When you connect Claude.ai or ChatGPT, the AI client sends tool calls to our server over HTTPS and we return tool results to that client. The AI client (Anthropic or OpenAI) sees the tool results, which are then included in the model's response to you according to that provider's data handling.

We do not send your data to any third party from our side. The AI client's privacy policy governs what they do with the tool results they receive.

Authentication and storage

The Connector authenticates via OAuth 2.1 with PKCE. Access tokens are short-lived (1 hour) JWTs signed with our private key. Refresh tokens (30 days) are stored as SHA-256 hashes; we never store the raw token. You can revoke any active session from your MCP settings.

We log connection events (token issuance, refresh, OAuth consent) for security and debugging. We do not log the contents of tool results.

Retention

Tokens and OAuth client registrations are deleted when they expire or when you revoke them. Audit logs are retained per the operational retention policy described in our main Privacy Policy.

Your controls

  • Revoke any connector at any time from your MCP settings or inside the AI client (Settings → Connectors → Disconnect).
  • Mint a public-only token (audienceIds: ["everyone"]) instead of the default self-bound token, so the Connector sees only data you've already marked public.
  • Audit your account's OAuth grants from your settings page.
  • Contact privacy@xhuman.com for a copy or deletion of your data.

Contact

privacy@xhuman.com