Prostir

Research-backed article

How to create a personal AI agent you own

How to create a personal AI agent you own: pick one job you repeat, add your sources and test it in your own AI app.

7 days free. No card required.

Before you read

What you get

A plan for one personal agent, its sources, memory, tools, access and tests, ready to use from several AI apps.

Who it's for

People setting up a private agent for one recurring part of their work or life, with clear limits on data, access and actions.

Topics

Personal AI agent · Knowledge and memory · MCP · OAuth

01

Start with one job you repeat

Pick one recurring job and describe it in Builder. Say what a useful result looks like, give it the context it needs and try it on real requests.

Many people are tired of every new chat asking the same questions, and of rebuilding their preferences each time they switch tools. Saving every conversation doesn't fix that.

Write down four things instead: what the agent should remember, what stays in a trusted source, what it must ask again and what it must never guess.

02

List your sources and actions

Start with a one-page list. Include the questions you ask often, approved documents, private facts, connected apps, what the agent may read, what it may propose to change, where it must ask you first and what gets deleted when.

The agent should answer from a source you own and can review. Give every lasting fact a source, a date and a way to correct it.

Where a calendar, repository or catalog already holds the data, that app stays the main record.

03

Keep knowledge, memory and tools apart

Use Knowledge for approved material that lasts, User Memory for small private preferences, and tools for live lookups or actions you allow.

A remembered preference can never grant access, publish, buy or delete anything. Changes to sensitive data need the real user, narrow permissions, a confirmation, history and a way to revoke access.

A first version may only search and explain. Add actions once you can see who acts, check the input, log the result and undo a mistake.

04

Use one agent in every AI app

Publish one agent and connect it to ChatGPT, Claude, Cursor, Codex or another app through MCP, the standard way AI apps connect to outside tools. Connect it only where that app and your account support MCP and sign-in.

Each app is only the window. The agent keeps the same instructions, knowledge, tools and access rules, so you don't rebuild it in every app.

The Prostir connection guide has the exact steps for ChatGPT and Claude.

05

Launch small and check the results

Keep the agent private with OAuth sign-in while you test, and test with a separate user account. Check current app support before you promise it works everywhere.

Try typical questions, missing or outdated data, a refused action and one recovery. Count finished jobs, corrections, safe refusals, time saved and upkeep.

Grow it only after the small version stays accurate, and drop data or tools that aren't worth their risk and upkeep.

This approach fits one recurring job you own. It doesn't aim at a general assistant that runs everything on its own.

Solutions

Build your agent from one prompt

Start with one customer question or routine task. Builder sets up the agent with your files and the actions you allow, and Prostir hosts and runs it.

Image preview