Personal AI Agent: What It Is and How to Run Your Own

2026-11-10

A personal AI agent is an AI system that acts for one person: it remembers them, takes goals and schedules, uses tools, and reports back where they already talk. POLYROB is an open-source personal AI agent you run on your own machine, with your own model, and this guide shows what it does and how to set one up.

The phrase is used loosely, so the first job is to pin it down. The second job is to show what a working one looks like in practice. Every POLYROB claim below comes from the public repository at release v1.7.6, and the sources are listed at the end.

What is a personal AI agent, and how is it different from a chat assistant?

Aembit’s glossary, checked on 2026-11-10, gives a clear working definition:

“A personal AI agent is an AI system that acts on behalf of one individual, learning their preferences over time and executing multi-step tasks across the services they connect.” — Aembit, Personal AI Agent glossary entry

The useful part is the contrast with a chat assistant. A chat assistant answers when you type and does nothing between sessions. A personal agent keeps going after you close the window.

	Chat assistant	Personal AI agent
When it works	Only while you type	Also between your messages, on a goal or a schedule
What it keeps	A session, sometimes a short history	Memory across sessions, an identity, a goal board
What it does	Answers	Plans steps, uses tools, follows up, reports back
Who holds the data	Usually the vendor	Depends on where it runs

That last row matters. A personal agent holds a lot about you, so where it runs is part of the product. POLYROB is self-hosted. There is no hosted POLYROB and no account to create.

What should a personal AI agent be able to do?

Five abilities separate an agent from a chat window. Each one below is a real, shipped feature of POLYROB, with the command you would use.

Does it remember you between sessions?

Cross-session memory is on by default and scoped to you. Out of the box it uses keyword recall over SQLite. An optional local vector index adds semantic recall. If the vector extra is missing, the agent logs one warning and falls back to keyword recall instead of failing. Run /memory in the terminal to inspect what the active provider holds. Memory and state are local SQLite files under ~/.polyrob/data. How recall ranks rows, how memory forgets and which settings to check are covered in AI agent memory.

It also keeps one conversation across devices. Every line you exchange on the terminal, the web console and your phone is recorded in one durable thread. A fresh session reads the tail of that thread on its first turn, so a reply to something said three minutes ago on another device lands with its context. /thread shows it.

Can you give it a personality you control?

Yes, in layers. The SOUL is a pair of plain Markdown files you write: who the agent is, its mission and its boundaries. The agent can never edit it. Below it sits a character, the voice, stored as a JSON file you pick with polyrob identity persona.

polyrob identity soul init
polyrob identity persona init mybot --from writer
polyrob identity persona show

persona show prints exactly what the model receives. Six curated presets ship with the install: analyst, coder, default, ops, researcher and writer.

Does it work on goals while you are away?

It can, once you switch that on. A goal board and scheduled jobs live in SQLite and survive restarts. A goal is one piece of work done once. A job is the same work at set times.

polyrob autonomy on
polyrob goals create "Draft the weekly report" -b "Summarize this week's notes in ./notes into report.md"
polyrob job new "every day 09:00" Summarise the inbox and post the three items that need me

Schedules accept a duration such as 30m, a phrase such as every monday 09:00, or a cron expression. This is off on a fresh install. polyrob autonomy on is the deliberate switch, and polyrob autonomy status shows every setting behind it. For the full picture of the levels and the risks, read the guide to the autonomous AI agent.

Does it live in the apps you already use?

POLYROB talks to you in the terminal, in a web console, and in eight chat surfaces: Telegram, email, WhatsApp, Discord, Slack, Signal, Feishu/Lark and DingTalk. polyrob gateway runs every enabled surface in one process. A new surface starts locked, so the agent answers no stranger until you let them in.

polyrob surfaces add telegram

Can it do real work with tools?

It reads the web with web_fetch (a URL becomes markdown, no browser needed) and drives a browser with Playwright. It works with files, git and GitHub, a persistent shell with background jobs, its own mailbox, and an MCP client. It can hand work to sub-agents that get a narrower toolset and cannot delegate again. Code execution is off by default. Reusable procedures it can load on demand are covered in the guide to agent skills.

How do you keep a personal AI agent under control?

An agent that knows you and can act for you needs limits you set first. POLYROB’s limits are mechanisms, not promises:

A step cap and a spend cap on every autonomous run. By default a run stops at 150 steps or $2.00 of provider spend and says why. This is a cap per run, not a daily or monthly budget, and the spend check can fail open if the probe errors.
A stop button that needs no model. /pause (alias /halt) is a slash command. It works on Telegram, in the REPL and in the console, and it survives a restart. polyrob autonomy pause does the same from a shell.
An answer to “what may it do right now”. /mode prints who you are talking to and what runs without asking. /mode ask-first, /mode standard and /mode on-my-own set one level. None of them touches money or access to your computer.
An approval queue. /pending lists everything waiting on you. A background run cannot wait for an answer, so a decision it needs becomes a durable ask that you answer from any device.
Untrusted input stays data. Web pages, mail and tool output reach the model as untrusted content, never as instructions. On a chat surface only the owner steers. This lowers prompt-injection risk. It does not remove it.
Self-edits wait for review. New skills and identity edits the agent writes sit in quarantine until you approve them.

Wallet and payment features exist as an opt-in extra and are off by default. The crypto code has had no independent audit. If you ever enable it, start on a testnet with a small cap.

How do you run your own personal AI agent?

You need Linux, macOS or Windows with WSL2, and one model: an API key, or a local model server such as Ollama. The installer fetches Python 3.11 or newer if you do not have it.

curl -fsSL https://polyrob.dev/install.sh | bash
polyrob doctor
polyrob

The installer clones the source, builds a virtualenv, puts polyrob on your PATH and starts the setup wizard, which asks for a provider key, a model, a persona and a chat surface. polyrob doctor is the preflight check. It names the one command that fixes anything missing. Then type a task in the REPL, or run one directly:

polyrob run "summarize https://example.com"

To keep it running after you close the terminal, install the background service:

polyrob service install
polyrob service status

On the model side, OpenAI, Anthropic, Google Gemini, DeepSeek, OpenRouter and NVIDIA NIM are built in. Any OpenAI-compatible or Anthropic-compatible endpoint, including a local one, works through ~/.polyrob/providers.yaml. /model <name> switches the model in the middle of a chat. For the full install walkthrough, with a hosted and a local model route, see how to self-host an AI agent and the install page.

One honest note on data. Memory, goals and sessions stay on your machine. Your prompts reach the model provider you chose, so a hosted model sees them and a local model does not. Since version 1.7.5 the installer sends one anonymous count after a fresh install (version, OS and install method, with no ID). --no-telemetry or DO_NOT_TRACK=1 turns it off.

Want more than one agent? polyrob profile create scout gives a second agent its own config, memory, skills and goals on the same machine, and polyrob -P scout runs it.

Where does POLYROB fall short as a personal AI agent?

Check these before you commit.

Fewer chat apps. It supports eight surfaces. Two other agent projects list about 29 and about 33 adapters, as of 2026-10-27 per the POLYROB comparison table.
No native apps and no wake-word voice. It runs from a terminal, a web console and chat apps.
You are the operator. You install it, update it and watch its cost. polyrob update --apply snapshots, migrates, verifies and rolls back on a git install, but a pip install only gets a printed command.
Local models are your job. POLYROB connects to a model server you run. It does not download or size a model for you, and a small local model plans and uses tools less reliably than a large hosted one.
A young project with a small community. The wallet code is unaudited, and tenant isolation in the multi-user posture is a software boundary.

If you are weighing it against other self-hosted agents, the OpenClaw alternatives guide and the compare page give dated, sourced rows.

How do you choose a personal AI agent?

Ask the same eight questions of any candidate, including this one.

Where does it run? On your hardware or on a vendor’s server.
What does it keep? Memory across sessions, and in what format.
Who writes its identity? You, or the vendor.
Does it work between messages? On goals and schedules, and does that survive a restart.
Where do you reach it? Which chat apps, and is one conversation shared across them.
What stops it? A step cap, a spend cap and a stop command that works without calling a model.
What waits for you? Do risky actions queue for approval, and can you answer from your phone.
Which models can it use? Can you bring a key, or a local model.

Personal AI agent questions

What is the difference between a personal AI agent and a chatbot?

A chatbot answers a message and stops. A personal AI agent keeps memory across sessions, works on goals and schedules between your messages, uses tools, and reports back. The difference is what happens when you stop typing.

How risky is a personal AI agent?

The risk depends on the limits you set. In POLYROB, autonomy is off until you turn it on, every autonomous run stops at a step cap and a spend cap, high-impact actions wait in an approval queue, and /pause stops background work. These reduce risk. None of them makes an agent risk-free, so start with small caps and low-impact tasks.

Can I run a personal AI agent without paying for an API?

Yes, with a local model. Declare your Ollama, LM Studio or vLLM endpoint in ~/.polyrob/providers.yaml. POLYROB itself is MIT licensed and free. You supply the hardware, and quality depends on the model.

Does a personal AI agent keep my data private?

Its memory, goals and sessions stay in files on your machine. Prompts go to your model provider, and the tools you switch on reach their own services. A local model keeps prompts on your machine.

Can I talk to it from my phone?

Yes. Add a chat surface such as Telegram with polyrob surfaces add telegram, and the same agent and the same memory answer there, in the console and in the terminal.

Where do you start?

Install it, run polyrob doctor, and give it one task before you switch on autonomy. The shortest path is two commands:

curl -fsSL https://polyrob.dev/install.sh | bash
polyrob doctor

Install POLYROB, read the docs, or see the full feature list on the capabilities page. The source is MIT licensed at github.com/theselfruleorg/polyrob.

Sources

All POLYROB facts are from the public repository at release v1.7.6 (github.com/theselfruleorg/polyrob).

README.md (what it does, safety, surfaces, quick start), docs/guide/instances.md (SOUL, character, presets), docs/guide/conversation.md (one thread across devices), docs/guide/streams.md (goals and jobs), docs/guide/owner-controls.md (/pause), docs/guide/configuration.md (memory, autonomy dial), docs/guide/profiles.md, docs/guide/chat-commands.md, docs/CONFIGURATION.md (step and spend caps), cli/ui/commands/h_mode.py (/mode levels), and docs/comparison.md (dated 2026-10-27).
Definition of a personal AI agent: Aembit glossary, fetched 2026-11-10.

POLYROB is an open-source AI agent you run and own. Install it or read the docs.