I Put an MCP Server on My Own Life Database

Most “AI in your productivity app” features are a text box wired to a completion endpoint. You get summarisation, maybe a rephrase button, and a settings page asking for an API key. It’s fine. It’s also not what makes an assistant useful, because the assistant can’t see anything you haven’t pasted into it.

The interesting version is the other way round: give the assistant a way to query and change your actual data, and let it work out what to do. That’s what the Model Context Protocol is for, and it’s why Life Manager Pro ships an MCP server with 39 tools over your tasks, projects, CRM, goals, habits and meetings — running locally, against the SQLite file on your own disk.

This post is about why that’s worth doing, what it actually enables, and the parts I’d want to know about before installing something like it.

The shape of the thing

Life Manager Pro keeps everything in one SQLite database on your Mac: contacts, companies, a deal pipeline, projects, tasks, notes, journal, goals, habits, meeting notes, and synced calendar events. No account, no server, no sync.

That turns out to be an unusually good substrate for an MCP server, for a boring reason: the data is already joined. A tool that answers “what’s overdue for this client” doesn’t need to reconcile three SaaS APIs with different ideas of what a person is. It’s a query.

The server is a sidecar binary the app ships with. It speaks MCP over stdio, so there’s no port to open and no endpoint to secure. Settings › MCP Server writes the Claude Desktop config for you, or hands you the binary path for any other client.

What 39 tools buys you

Roughly three layers.

Generic access to everything. list_tables and describe_table let a client discover the schema at runtime, and create_entity / get_entity / update_entity / delete_entity / list_entities operate on any table with a filter tree, sorting and pagination. Plus search, which is full-text across tasks, projects, contacts, notes, goals, deals, meetings and journal entries.

This layer means the server doesn’t need a bespoke tool for every question. An assistant that can read the schema and run a filtered query can answer things I never anticipated.

Purpose-built tools for the things worth doing well. Generic CRUD is flexible but it makes the model do work it’s bad at — multi-table joins, date arithmetic, ranking. So the common questions get real implementations: get_daily_focus, get_overdue_tasks, get_dashboard_stats, get_project_health, get_active_goals_summary, get_executive_overview, get_recent_activity, list_contact_deals, list_interactions, list_key_results.

The one I use most is get_meeting_dossier. It takes a meeting and returns the counterpart contact, their company, active projects, decisions from past meetings, open deliverables and live CRM deals — the same briefing the app renders in its pre-meeting panel. Asking “brief me on my 2pm” and getting a real answer is the single feature that changed how I use the thing.

Writes. This is the part that makes it more than a search box. complete_task, complete_project, log_habit, update_key_result_progress, create_task_from_action_item, add_tag, link_entities, plus generic create/update/delete.

“Turn yesterday’s meeting action items into tasks assigned to the right people” is one instruction, and it lands as real rows linked back to the action items they came from.

The privacy argument, stated honestly

The obvious objection to letting an AI at your client list is that you’ve just handed your client list to an AI company. Worth being precise about what is and isn’t true here.

The server sends nothing anywhere. It’s a local process reading a local file, and it works with the network off.

What does leave is whatever the assistant reads while answering you, through whichever client you connected — so if you ask Claude Desktop about a deal, the relevant rows go to Anthropic as part of that conversation, the same as if you’d typed them.

The difference from a cloud tool is where your data sits by default. With a hosted notetaker or CRM, your records are on someone’s servers before any AI is involved, and the vendor decides what’s queryable. Here the default is your disk, the exposure is per-question, and if you don’t want any of it you simply don’t install the client integration — the app is fully functional without it.

That’s a real distinction, but it’s a distinction about defaults and control, not a claim that AI is involved without data moving. Anyone telling you otherwise is selling something.

Where the sharp edges are

Three things I’d want to know.

It can change things. An assistant with write access will occasionally do something you didn’t quite mean — close the wrong task, tag the wrong record. Most deletes here are soft, so Trash is a real undo, but “read-only would be safer” is a legitimate preference and this isn’t that.

Generic CRUD is a big surface. Exposing every table to a model is powerful and slightly alarming. The mitigation is that it’s your machine, your file, your client, and nothing is exposed to a network — but it’s a deliberate trade, not an oversight.

Quality depends on what you’ve filed. The dossier tool is only as good as your CRM. If you don’t record decisions, it can’t surface them. MCP doesn’t fix an empty database.

Why this belongs in a personal app

There’s a version of this argument that says agents will make personal productivity software obsolete — why keep a task app when an assistant can just manage things for you?

I think it’s the opposite. An assistant needs somewhere durable to put state, and a chat log isn’t it. What it needs is a structured store with a stable schema, sensible defaults about what a project is, and a UI for the human to review and correct what happened. That’s what an app like this is. MCP is the seam where the two meet.

The bet is that the interesting personal software of the next few years isn’t AI-native replacements for your task manager, but ordinary apps with a good local data model and an honest interface for an agent to drive them.

If you want the full tool list and setup, it’s on the MCP server page.