How to connect a personal AI agent to your network

Connect the data, not the files

The fastest way to give an agent your network is to paste in an export. The way that keeps working is to connect the agent to a system that already holds your network and keeps it current, so the agent can ask questions and write back instead of rereading files. Why the second approach holds up when the first does not is covered in can a personal AI agent run your CRM on its own.

The standard for that connection is MCP, the Model Context Protocol. A service that holds your data runs an MCP server offering tools such as search contacts, show how I know this person, add a note or draft an introduction. Any assistant that supports MCP can call them. You set up the connection once per assistant and the data stays in one place.

Before you start

You need two things.

A system that holds your network. It should sync your email, calendar, contacts and LinkedIn, merge duplicates, and keep notes and history, because the agent will only be as good as what it can query. The examples below use NetworkOS, whose server is at https://www.thenetworkos.com/api/mcp. Any personal CRM with an MCP server follows the same pattern; Dex, for example, offers an AI skill for Claude Code.

An assistant that supports MCP connectors. Claude, ChatGPT on a paid plan, Meta Muse and most agent frameworks do.

In Claude

1. Open Claude's connector settings (Customize, then Connectors) and choose to add a custom connector.
2. Paste the server URL and connect.
3. Sign in to the service when the consent screen opens and approve the access it asks for.
4. Start a new conversation and ask something only your network can answer.

The connector is then available in every conversation. Claude shows interactive cards for some results, such as a contact list to pick from or an introduction to review before it sends.

In ChatGPT

ChatGPT runs full MCP connectors in Developer Mode, which is available on paid plans.

1. Turn on Developer Mode in ChatGPT's settings.
2. Create a new connector, give it a name, and paste the server URL as the MCP server address.
3. Sign in when ChatGPT prompts you.
4. In a chat, open the tools menu, go to Developer Mode and switch the connector on.

The same server, tools and cards work here as in Claude, because both hosts build on MCP.

In agents that take a token

Some agents do not run a sign-in flow and instead accept a bearer token. Meta Muse is one: you ask it to build a custom integration to the server's URL, tell it the server is remote over HTTP with bearer token auth, and paste a token into its credential prompt rather than the chat.

For that case NetworkOS issues a personal access token in Settings, under AI connectors, where you choose read-only or read and write before generating it. The token works with any client that sends a bearer token, not only Muse. It is shown once, lasts a year and can be revoked there at any time. There is one live token per account, so generating a new one switches off the old one; if two agents need a token, they share it.

In your own agent or framework

Agents and frameworks that support remote MCP servers with OAuth sign-in connect the same way Claude does. Point them at the server URL; the server publishes its sign-in details at the standard discovery addresses, lets a client register itself, and accepts a local callback address such as http://localhost, so a command-line agent can complete sign-in in your browser. Clients that cannot run a sign-in flow use the personal access token instead. The full technical notes for agent builders, including the tool list and rate limits, are in the site's llms-full.txt.

Three tests to run first

Once connected, check that the agent is really using your network and not guessing.

1. A search. Ask who you know at a specific company. The answer should name people from your own contacts, ideally with a reason each one matched.
2. A lookup. Ask how you know one of those people and when you last met or wrote. The answer should come from your real meeting history, and from email history if you allowed the connector to read your mailbox, not from a guess.
3. A draft. Ask for an introduction between two people you know. The agent should produce a draft and stop for your review. Nothing should send until you say so.

If any of these comes back generic, the agent is answering from its own knowledge and the connector is not switched on for that conversation. If answers are thin rather than generic, the first sync may still be running; NetworkOS's get_sync_status tool tells the agent whether it has finished.

Permissions worth setting

Start read-only if the option exists, and widen access once you trust the answers. With NetworkOS that means a read-only personal access token, because Claude and ChatGPT ask for read and write when they sign in. Most hosts also let you require approval before each tool call; turn that on for writes while you are getting started. A good connector separates reading from writing and asks before each send, so a misunderstood instruction produces a draft, not a message in someone's inbox.

In NetworkOS, read tools run on their own, other changes such as notes and tags happen when the agent calls them, and every send waits for your explicit go-ahead; in Claude and ChatGPT the introduction appears in a card you can edit first. Archiving a contact can be reversed in the web app; deleting a note and sending an email cannot. The limit worth knowing is that interactive cards depend on the host: Meta Muse shows results as text, and it uses its own approval prompt before a send.

What the agent can do day to day once it is connected, prompt by prompt, is in what you can do with your network inside Claude and ChatGPT.