What is Agent City?
Agent City is a free Mac app that gives you a team of small AI agents. Each agent does one job: telling the time, checking the weather, converting currency, searching the web, working with your files. You chat with the whole team at once, and Agent City works out who should answer.
It runs on your computer, on a language model you choose. That can be a free model running on your Mac through Ollama, or a cloud model from OpenAI, Anthropic or any OpenAI-compatible provider.
Built for three things
Section titled “Built for three things”1. Build the agents you need
Section titled “1. Build the agents you need”Describe an agent in plain English: “an agent that reads my Home Assistant and tells me which lights are on”. The Agent builder uses Claude Code to write it, test it and add it to your team in minutes. Each agent is a small folder of plain files you own.
2. Put them to work on their own
Section titled “2. Put them to work on their own”Automations run your agents on a schedule, when a device sends an event, when a message arrives, or when a call starts. Describe the routine and Agent City drafts the steps: gather, check, ask you first, summarise, report.
3. Run it all on your own hardware
Section titled “3. Run it all on your own hardware”Once your agents are built, everything can run completely on-premises: a local model through Ollama, local voice, local storage, and your own network. No cloud model, no Claude Code, no account. See Run it all on your Mac.
How it works
Section titled “How it works”You talk to the coordinator
Section titled “You talk to the coordinator”The main chat, My team, is run by a coordinator. It has no tools of its own. When you ask something, it:
- decides which agents can help,
- writes each of them a complete question,
- asks them (several at once when it can), and
- combines their answers into one reply.
You can watch this happen. Each reply shows the agents the coordinator asked and the tools they used, and you can expand that view to see the details.
Follow-ups like “And in Tokyo?” work because the coordinator remembers the conversation and fills in what the agents need to know.
Agents are capability packs
Section titled “Agents are capability packs”An agent doesn’t act on its own. It’s a set of tools with notes on how to use them, plus any keys or sign-ins those tools need. The coordinator does the thinking and planning. This keeps each agent small enough for a local model to use reliably.
Agents can also ask each other for help. A question can go from the coordinator to one agent and then on to a second, but no further, so questions can’t bounce around forever.
Memory everyone shares
Section titled “Memory everyone shares”When you tell Agent City something lasting about yourself, like your name, your home city or the units you prefer, it saves it to Shared memory. Every agent sees those facts, so you never have to say them twice. You can view, edit and delete them at any time. See Memory.
What’s included
Section titled “What’s included”| Kind | Examples | Notes |
|---|---|---|
| Starter team | Tock (time), Nimbus (weather), Penny (currency), Scout (web search) | Ready straight away. No accounts or API keys needed. |
| Built-in agents | System, Computer, Meetings | Part of the app. They can’t be removed. Computer and Meetings are off until you turn them on. |
| Messaging agents | Telegram, Signal, X | Read your own accounts once you connect them on Connections. |
| Installed agents | Flights, shares, news, Gmail and more | Installed from agent sources on GitHub. |
| Your agents | Anything you need | Build them with the Agent builder, or add an MCP server. |