Personal MCP server for AI context and preferences

Bring your preferences to an AI tool

Project Fingerprint hosts a remote Model Context Protocol (MCP) server for your personal AI profile and taste map. A compatible AI client can read an approved copy of selected profile sections, likes, interests, and practical preferences. You control each connection’s access; your full private vault stays separate.

Connect in five steps

1. Write the context you want to reuse in About me. Sign in to your Fingerprint account and unlock it.
2. Open Connect AI → Add connection and choose API / MCP. Select profile sections and, optionally, taste map data and suggestion permissions.
3. Select Preview sharing and review the exact data. Continue, name the connection, leave Authorization set to OAuth, and select Create connection.
4. Add the server address to your AI client. During its OAuth flow, sign in to Fingerprint if asked, select the connection you created, and approve the requested access.
5. Ask the client to read your approved context. Creating a connection alone does not confirm that the client has used it.

Approve the access your client needs

Initial discovery requests ordinary profile and taste reads. Fingerprint lists the scopes allowed by each connection before approval and issues only that subset. Your client must check the returned scope list. Proposals, saves, and candidate comparisons require their separate requested permissions.

A known tool that needs a missing token scope returns an HTTP 403 permission challenge. Additional access requires another consent flow. If the owner approves only part of the request, the client must respect that subset. Token scopes cannot override the connection’s sharing restrictions.

Client registration

Compatible public clients can identify themselves with an HTTPS client metadata document or use dynamic registration. Both use PKCE S256 and an exact registered callback. Fingerprint refreshes URL-based client details when you review and approve a connection; changed details require another review. Check the client ID shown on the approval screen.

Developers can find the metadata format and discovery endpoints in the API integration guide.

Try a request

After sharing Food and Practical preferences, ask: “Use my approved Project Fingerprint context to suggest three dinner options. Explain which shared preferences support each option, ask about missing constraints, and do not change my profile.”

With profile suggestion permission, you can also ask the client to interview you and send one profile draft for review in Fingerprint.

Keep control of access and updates

The server can read the last approved shared copy while Fingerprint is closed. That copy may be old. Open Fingerprint unlocked and online to refresh it; context responses include update times.

Suggested changes need review by default. You can separately permit a remote connection to save selected profile sections or taste ratings. Authorized saves wait until Fingerprint is unlocked and online, and conflicts stop for review. A queued update is not yet saved. New sections remain private until you approve sharing them.

Review, pause, or revoke a connection in Connect AI. Revoking access cannot recall copies a client has already received. Read the AI sharing controls for details.

API, MCP, browser tools, or Markdown?

• JSON API: apps and backend services read approved context with HTTP requests and JSON responses.
• Remote MCP: a client connects to the hosted server through an approved account connection.
• Browser WebMCP: a supported browser agent uses tools in your open Fingerprint tab. Sharing approval ends on reload, sign-out, account change, or Stop sharing.
• Markdown: export aboutme.md from your local profile and give the snapshot to a tool that accepts text or files. No account is required to write or export. Later edits require another export.

These options provide context to tools; they do not automatically synchronize an assistant’s built-in memory.

For integration developers

The remote get_profile_template tool returns public starter headings, not personal data, but calls to the remote MCP endpoint still require authentication. Reading selected personal sections through get_profile_context also requires profile read permission. Browser WebMCP can expose the public template before the user approves personal reads.

To use HTTP and JSON directly, follow the API setup guide and OpenAPI specification. The API uses the same OAuth connection and permissions as remote MCP.