AI agents as relationship copilots: useful, careful, ours

An LLM with read access to your contact log is genuinely useful for narrow tasks. Find me everyone overdue I haven't talked to in three months. Draft a message for Sara's birthday tomorrow. Group the people I met at the conference last November. These are queries you could answer manually - the agent just makes them ten seconds long.

Contact Book ships an MCP server (Model Context Protocol) that exposes the data-model side of the product to any compatible client - Claude Desktop, ChatGPT-with-MCP, your own scripts. The MCP surface is intentionally narrow: read + write rows, no auth or billing tools. We treat that boundary as load-bearing.

Three things to be wary of. First, autonomous outreach. An agent sending messages on your behalf sounds great until the first awkward draft lands in someone's inbox at 2am. Keep the human in the loop on send. Second, summarising relationships. Asking an LLM to "summarise my friendship with X" produces fluent prose that's also subtly wrong, and you won't notice until you greet someone with a fact you fabricated. Third, learning patterns across users. We don't train on your data, period.

The agent surface uses the same auth + access gates as the regular API. There's no special "agent" token that bypasses anything. Rate limits, audit logs, soft-delete, and the WebSocket fan-out invariants all hold. If your agent does something destructive, it shows up in the audit log just like any other action you took.

If you stop using the agent, you revoke the personal access token. The data is unchanged. We don't keep agent traces, we don't profile your behaviour, we don't create a model fingerprint of you. The agent is a tool that uses the API; it doesn't get a separate identity.