Skip to content
The AI Agent Factory

14.9 How today's tools realize it

Approved by
Panaversity ·
8 min read

In everyday life. Three office buildings all take deliveries. In the first, the front desk calls you before any parcel comes up. In the second, the doors stay locked, and you let each driver in yourself. In the third, drivers walk straight in unless you post your own rule at the door. In every building, you read the label before you sign for the parcel.

What you will learn. Where Claude Code, Codex and OpenCode differ when they install a package, reach a server on your own computer, keep it running and review a change, as of October 30, 2026.

Why it matters. Claude Code is the front desk: it asks before each install. Codex is the locked door: its sandbox keeps the network off until you approve. OpenCode is the open door, until you set its rules. The parcel is the package. Your part does not change: you decide, you predict, the agent runs it, and you read what really happened.

How to use this page. This page holds only what differs by tool. Pick your tool once, and every strip on this page follows.

How does each agent ask before it installs a package?

In these tabs, a sandbox is not the throwaway machine of 12.3. It is a set of limits on your own computer.

An install, such as npm ci, npm install zod or uv add pydantic, is a shell command, so Claude Code asks before each one in Manual and Accept edits modes.1 Read the package name in the prompt before you approve. Use Manual for this page.

In the workspace-write sandbox, network access is off by default, and Codex asks for approval before a command that needs the network, such as npm install.2 So expect it to ask before npm ci and uv sync, which download packages. Use Ask for approval for this chapter's Try it steps.

OpenCode runs npm and uv with your own rights.3 Before your first message in a folder, set edits and the shell action to ask, as in Lab 12, so each command and each install waits for you.4 OpenCode does not sandbox the agent.5

How does each agent reach a server on your own computer?

Whichever agent you use, a server it starts is stopped when the task ends.

A local session has no network limit by default. With the sandbox on, a local test server needs an exception.6 Claude Code can run a long-running command, such as a development server, in the background while it keeps working. /tasks lists background tasks.7

OpenAI's security page says that when its network proxy is on, local and private addresses such as localhost are blocked unless you allow them. The proxy is off by default.2 So check here whether your Codex session can reach the sample API, before the lab needs it, and allow it in your settings if it cannot. Codex can run commands in background terminals. /ps lists them with their recent output, and /stop stops them.8

A shell command has your full file, process and network authority, so it can reach localhost and the internet alike.3 The shell tool can start a command in the background.3 Before the task ends, check the process list and stop what is still running. The key here is made up. With a real key, remember that OpenCode can read any file your account can.

How does each agent review a change?

Run /code-review before the commit.9

/review reports findings on your uncommitted changes or your branch, without changing files.10

OpenCode's version 2 documentation describes no review command. Read its review panel or git diff instead.11

What it costs

Claude Code comes with the Pro, Max and Ultra plans,12 and with Team and Enterprise seats that include it, as of November 1, 2026.13

Codex is on every ChatGPT plan, including Free.14

OpenCode is free under the MIT license, at version 2.2.5 on October 30, 2026.15 You pay for the model you connect.

Check yourself

Question 1 / 6

0 answered

In Claude Code's *Manual* mode, what happens when the agent wants to run npm install zod?

  1. Permission modes, Anthropic.↑
  2. Agent approvals and security, OpenAI.↑↑2
  3. Tools, OpenCode.↑↑2↑3
Owner
Panaversity