Run it all on your Mac

Agent City can run completely on-premises. Once your agents are built, nothing needs to leave your Mac: no cloud model, no Claude Code, no Agent City account.

Build once, run locally

  Build (once, online)                         Run (every day, on your Mac)
  ─────────────────────                        ────────────────────────────
  Agent builder + Claude Code   ──writes──▶    agents/recipes/SOUL.md
                                               agents/recipes/tools.py
                                                        │
                                                        ▼
                                               Ollama · qwen3.5:9b (local)
                                               chats · automations · events
                                               memory · reminders · meetings

1. Build your agents. The Agent builder uses Claude Code, with your Claude account, to write each agent’s files.
2. That’s the last time Claude is involved. An agent is just a folder of plain files (SOUL.md and tools.py) in your data folder. Running it needs only Agent City and your model.
3. Run everything on a local model. Point Agent City at Ollama (qwen3.5:9b is the default), and every chat, automation, event and reminder is handled on your Mac.

You can also skip the builder entirely and write agents by hand, or install them from a source.

What runs locally

Part	Runs on your Mac with
Every agent’s thinking	Ollama, LM Studio, or any OpenAI-compatible server on your Mac or network
Chats, memory, history, files	SQLite files in ~/Library/Application Support/Agent City
Automations and schedules	Agent City’s own scheduler
Incoming events (the Ingest API)	Agent City’s own server, on your Mac or your network
Voice typing	The built-in Whisper
Reading aloud	The built-in Kokoro
Meeting transcripts and notes	Whisper, plus your local model
Drafting automations	Your agents’ own model (the default)

What can still reach the internet

Agent City itself doesn’t need the internet. These things do, and each one is your choice:

What	When	To stay offline
Agents that use online services	When they’re asked something, like the weather, web search or Gmail	Use agents whose tools only work with local data, and turn the others off on the Agents page.
The System agent’s web tools	Only if Read web pages is on	It’s off by default. Leave it off.
Cloud models	Only if you choose one	Use Ollama or another local server.
MCP servers	Remote servers (url) are reached over the internet	Use servers that run locally (command).
Agent sources	When you browse or refresh them	Don’t refresh. Installed agents keep working without their source.
Claude Code	Building agents, and drafting automations only if you choose Claude Code for it	Build first, then you don’t need it.
Update checks	Shortly after launch, then every 6 hours	Updates are only installed when you choose. The check itself can’t currently be turned off.

A fully local setup

1. Install Ollama and let the Setup check download qwen3.5:9b.
2. In Settings → Models → Model, choose Ollama on this computer.
3. In Settings → Models → Drafting automations, keep your agents’ own model.
4. Install Whisper and Kokoro from Settings → Models if you want voice.
5. In Settings → Tools → System tools, leave Read web pages off.
6. On the Agents page, turn off any agent that talks to an online service you don’t want to use (each card says what it uses).
7. Build the agents you need, or write them by hand, so their tools only reach things on your Mac or your network: your files, your home automation server, your NAS, a local database.

Your home network counts as local. An agent’s tools can call your own servers (Home Assistant, a Raspberry Pi, an internal API), and devices on your network can send events to Agent City.

Keeping it on-prem

• Keys and data stay on your Mac. API keys, sign-ins, chats and memory are stored in your data folder, never sent to Agent City’s makers. See Privacy and your data.
• Use more powerful local hardware. A local model doesn’t have to be on the same Mac. Point Agent City at an Ollama or vLLM server elsewhere on your network (in Settings → Models, choose another OpenAI-compatible server and give its address). Agent City treats addresses on your network, like 192.168.…, *.local or Tailscale, as local.
• Check where requests go. Settings → Data & privacy → Where your data goes summarises it for your current setup, and the Audit log records every tool call any agent made.