Agent Zero · main guide

Step 8: Advanced Agent Zero

This page is for you once the playbook jobs feel routine and Agent Zero has run well for a week or two. It covers six power features, and each one comes with what it is for, how to turn it on, and what you trade for it. Each feature is optional, so try one at a time and keep everything else on this page as it was.

Applies to: Agent Zero only. The setups that work with any agent are listed at the end.

Everything here happens on the Agent Zero web page at http://localhost:5080 or in A0 Launcher, and the clicks are the same on Windows and on a Mac. These steps were checked on 14 November 2026.

Model presets: a cheap model for most work, a strong one on demand

What it is for. A model preset is a named set of models. The Main model talks and uses tools, and the Utility model writes summaries, chat names, and memory notes. The Embedding model searches memory. Presets let you switch the whole set from the chat in one click. A fresh install comes with presets called Default, Efficiency, and Power, and the preset menu is the first dropdown on the left of the chat status bar.

  1. Click the current preset name, then click Edit Presets.
  2. Click Add, name it Cheap, choose a low-cost model for the Main and Utility models, and click Save. For that low-cost model, the API id is deepseek-flash, and the version is DeepSeek-V4.1-Flash-1017.
  3. Click Add again, name it Strong, choose a frontier model for the Main model, and click Save.
  4. Open Settings → Agent → Models and make Cheap the default for new chats.

What you should see new chats start on Cheap. When a job needs careful planning, you pick Strong from the preset menu for that chat only.

A project or an agent profile can also have its own preset, through the Per-project / agent option in the same settings. Leave the Embedding model alone, because changing it means rebuilding the memory index. The prices behind this split are on Mix a cheap model with a strong one.

A local model on your own computer

What it is for. With a local model in Ollama or LM Studio, your prompts never leave your computer and there is no bill per step. Agent Zero runs inside a container, so it reaches your computer through a special address, host.docker.internal, rather than localhost. Inside the container, localhost means the container itself.

If you picked Local model during first-run setup in Step 1, Agent Zero already uses Ollama. The steps below add a separate Local preset, so you can switch between a cloud model and a local one from the preset menu.

  1. Install Ollama with Run a local model with Ollama, download a model in its terminal, and leave Ollama running. For example:

    ollama pull qwen3.8:27b

  2. In Agent Zero, click Edit Presets, click Add, and name the preset Local.

  3. For the Main model, choose Ollama and type the model name exactly as Ollama lists it, such as qwen3.8:27b. If Agent Zero asks for an address, use http://host.docker.internal:11434. For LM Studio, the port is 1234 instead.

  4. Click Save and pick Local from the preset menu.

  5. If a small model explains what it would do instead of doing it, pick the Tiny Local profile from the profile menu beside the chat input. That profile is built for small local models.

What you should see a short test question gets an answer from the Local preset. If Agent Zero cannot connect, the model server may only accept connections from inside your computer. Local models covers that fix for Ollama and LM Studio.

Agent Zero's docs say the model server must listen on an address Docker can reach, and that you should "keep that port limited to trusted clients". They also warn that a Utility model that loads is not proven to handle summaries and memory well, so test it on a few chats before you rely on it.

Projects, agent profiles, and memory cleanup

What it is for. A project gives one kind of work its own files, instructions, memory, secrets, and model settings, so a client job never sees your personal notes. An agent profile sets a role and turns tools on or off. Memory is what Agent Zero has learned, and the Memory page lets you review it.

To make a project:

  1. Open Projects from the dashboard and click Create project.
  2. Give it a title, a short description, and plain instructions. Agent Zero's own example includes "Keep generated files inside this project unless I ask otherwise".
  3. Save it, open a chat, and choose the project from the project picker in the top right.

What you should see the chat shows the project as active, and files the agent makes land in that project.

To make a careful profile:

  1. Open the profile menu beside the chat input and click Manage agents, then Create agent.
  2. Set each tool to On, Default, or Off. A cautious start turns off the tool that runs terminal commands and code, and the text editor's Write and Patch actions.
  3. Save the profile and pick it for chats that only need to read and summarize.

What you should see the new profile appears in the profile menu, and its tool settings show those tools as Off.

Agent Zero's docs add a limit: "Tool permissions control those tool calls; they are not a filesystem sandbox". The same docs say an agent with terminal access can still write files even when the editor's Write action is off, which is why step 2 turns the terminal off too.

To clean up memory, back up first with Settings → Backup & Restore. Then open Memory from the dashboard, search for old test entries or facts that are no longer true, and edit or delete them. The docs say "A stale or incorrect memory can poison the processing instead of helping it," so this is the first place to look when the agent keeps repeating a mistake.

What you should see the entries you deleted no longer appear when you search Memory again.

Add-on servers and plugins

What it is for. MCP servers and plugins add new tools, such as control of a browser or a link to another app. Agent Zero lists them on the MCP/A2A tab in Settings and in the Plugin Hub.

To add an MCP server:

  1. Open Settings, click the MCP/A2A tab, find External MCP Servers, and click Open.

  2. Paste the server's settings in JSON. This example from Agent Zero's docs adds a browser debugging tool:

    {
  "mcpServers": {
    "chrome-devtools": {
      "command": "npx",
      "args": ["-y", "chrome-devtools-mcp@latest"]
    }
  }
}

  3. Click Apply now.

  4. To limit which profiles can use it, open the profile with Edit agent, go to MCPs, and set the server On or Off.

What you should see the server shows a green status and a count of its tools. A server that runs on your own computer, rather than in the container, needs host.docker.internal in its address.

For plugins, open Plugins and use Browse or Install to reach the Plugin Hub. Agent Zero's docs ask you to read each plugin's "description, README, permissions, and source link" before you install it, and to back up before plugin experiments.

Host access, done carefully

What it is for. Host access lets Agent Zero use files, commands, a browser, or Computer Use on your real computer instead of only inside its container. It starts off, and every earlier page keeps it off. The folder you pick sets where work starts, but Agent Zero's docs say it does not restrict commands to that folder.

Use a second instance with its own workspace for this, so your everyday Agent Zero is never the one with Host access.

  1. Back up your everyday instance first, using Backup /a0/usr on its card under Instances in Launcher.
  2. Make a new folder on your computer, such as agent-work, and copy in only the files for the job.
  3. In Launcher, open Installs, click Run, open Advanced, and under Workspace storage choose Create folder named after Instance. Agent Zero's docs say to "Use a separate workspace for each running Instance".
  4. Open the new instance with Open UI, click the computer icon beside its name in the Launcher tab bar, choose the access you need and the agent-work folder, and click Save.
  5. Leave Browser and Computer Use off unless the job needs them.
  6. When the job is done, close that instance's tab, which stops Host access.

What you should see the new instance can list the files in agent-work, and your everyday instance still says Host access is off.

The Flare tunnel, behind a login

What it is for. The Flare tunnel gives your Agent Zero web page a public web address, so you can use it from your phone or another place. Agent Zero's docs warn that "anyone with the tunnel URL can try to open your instance".

  1. Open Settings → Authentication and set a UI Login and a strong UI Password. If you installed with a login and no password, the installer's default password is 12345678, so change it now.
  2. Click Save, then open the page in a private browser window to check that it asks you to log in.
  3. Open Settings → External Services → Flare Tunnel and click Create Tunnel.
  4. Copy the web address it shows and open it on your phone.
  5. When you are done, stop the tunnel on the same screen.

What you should see the address starts with https://, and it shows the Agent Zero login page before anything else.

Agent Zero's docs put it plainly: "Always set a username and password in Settings -> Authentication before creating a tunnel".

Shared advanced topics

These pages cover setups that work with any agent in this guide. Each one applies to Agent Zero.

Sources

Checked 14 November 2026.

Next

Next: Persona and rules.

Checked 14 November 2026 · prices, limits, and versions change