Skip to content
Agent CityAgent City

Requirements

At a glance

Section titled “At a glance”
You need For Required?
A Mac with Apple silicon, on macOS 12 or later Running Agent City Yes
At least one language model that supports tool calling Every agent, every chat, every automation Yes
Claude Code, installed and signed in The Agent builder Only to build agents
Whisper (built in, one click) Voice typing and Meetings Optional
Kokoro (built in, one click) Reading replies aloud Optional

Your Mac

Section titled “Your Mac”
  • Apple silicon (M1 or later). The download is an arm64 build.
  • macOS 12 Monterey or later. Telling speakers apart in meeting notes needs macOS 15.
  • Memory: 16 GB or more is recommended for running a local model like qwen3.5:9b alongside your other apps. On an 8 GB Mac, use the smaller qwen3.5:4b, or a cloud model.
  • Disk: about 7 GB for qwen3.5:9b, plus about 700 MB for Whisper and 500 MB for Kokoro if you add them.

A language model

Section titled “A language model”

Every agent thinks with a language model, so you need at least one. It must support tool calling (also called function calling). Agent City talks to it through an OpenAI-compatible API, so almost any modern model works.

Section titled “Recommended: Ollama and Qwen 3.5 9B”

The easiest start is free and runs entirely on your Mac:

  1. Install Ollama.
  2. Open Agent City. The Setup check finds Ollama, starts it if needed, and downloads the default model, with a progress bar.

The default is qwen3.5:9b. It’s the model Agent City is tuned and tested against: reliable at picking and using tools, quick enough on Apple silicon, and private.

Model Download Good for
qwen3.5:9b 6.6 GB Recommended. The default, and the best balance of speed and reliability on a 16 GB Mac.
qwen3.5:4b 3.4 GB 8 GB Macs, or when speed matters more than accuracy.
qwen3.5:27b and up Larger Macs with plenty of memory, and long background research.

To download a model yourself instead, run this in Terminal:

Terminal window
ollama pull qwen3.5:9b

Keep Thinking on Low for Qwen models. With it off, small models sometimes answer without using their tools. See Choose a model.

Or a cloud model

Section titled “Or a cloud model”

You can use OpenAI, Anthropic, OpenRouter, or any OpenAI-compatible server instead, or as well, with an API key from that provider. You can also mix them: a local model for most agents and a cloud model for the ones that need more power. See Choose a model.

Claude Code, for the Agent builder

Section titled “Claude Code, for the Agent builder”

The Agent builder uses Claude Code to write new agents for you. You need:

  • Claude Code installed (setup guide), so the claude command works in Terminal, and
  • Claude Code signed in with your own Claude account. The builder uses your Claude Code account and settings.

The Setup check shows whether Claude Code is found.

Claude Code is only needed while you build or change agents. Once an agent exists, it runs on your own model like any other, with no Claude Code involved. You can also use Claude Code to draft automations, but that’s optional: by default your own model drafts them. See Run it all on your Mac.

Optional extras

Section titled “Optional extras”
Extra Needs Used for
Tailscale A Tailscale account on your Mac and phone Using Agent City from your phone away from home. On the same Wi-Fi, nothing extra is needed.
Homebrew Installing signal-cli for the Signal agent. Agent City does the install.
API keys Each service’s own sign-up Agents for paid or private services. Each agent’s card says what it needs.