13.11 How today's tools realize it
- Approved by
- Panaversity ·
In everyday life. Three driving schools teach the same rules of the road in different cars. Two cars come with a brake on the instructor's side, and in the third the instructor fits one before the first lesson. You are the instructor in every car.
What you will learn. Where Claude Code, Codex and OpenCode differ when they run a command, reach the network, plan a change and review it, as of October 30, 2026.
Why it matters. The products differ in what runs without asking. Your part does not: you guess, 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 acts?
In these tabs, a sandbox is not the throwaway machine of 12.3. It is a set of limits on your own computer, like a contractor who works in your house but may change only one room, and asks before going out for supplies.
- In Manual mode Claude Code asks before every shell command, a command run in a terminal, except a built-in set of read-only ones.1 Use Manual for this page.
What happens when a tool comes from the internet?
The two type checkers come from the internet the first time, through npx and uvx. This is where the three products differ most:
A local session has no network limit by default, so the agent can install a package once you approve the command. The optional sandbox, turned on with /sandbox, routes commands through a proxy, a checkpoint for network traffic, that allows only the domains, or web addresses, you list.6
Codex asks for approval before a command that needs the network, such as npm install or uv add, which installs a Python package. Tests that need no network, such as npm test, run inside the sandbox.4 So expect it to ask before the first npx and uvx.
OpenCode does not sandbox the agent, and its permission prompts are not built to isolate it. Its security policy says to run it in a container or virtual machine if you need isolation.7 With your opencode.jsonc set to ask, it asks before each command, and each command runs with your own rights.
How does each agent plan without editing?
Plan mode reads files and proposes a plan, and edits nothing until you approve. Switch to it with Shift+Tab, or start a prompt with /plan.1
/plan asks for a plan first.8
OpenCode's plan agent explores and plans without editing project files, and the build agent edits and runs commands.9
How does each agent review a change?
/code-review reviews your branch's commits and your uncommitted changes for bugs and reports findings, without changing files unless you ask it to fix them.10
/review reviews your uncommitted changes, or your branch against a base branch, and reports findings by priority without changing your files.11
The differences that change a decision
- Who can install a package without a prompt. OpenCode can, if your rules do not say ask. Claude Code reaches the network freely once you approve a command. Codex's sandbox blocks the network by default and asks. On your laptop, approve each install only after you have read the package name.
- Where tests run. Codex runs commands on your computer, inside a sandbox that limits writes to the project. Claude Code runs them with your full rights unless you turn on its sandbox. OpenCode always runs them with your rights. So read each command before you approve it: outside a sandbox, it can change any file you can.
- Review built in. Claude Code and Codex have review commands, and OpenCode's Desktop has a review panel. On every product, an agent's review is a second worker checking the first. Chapter 6 says that is not the review, so the review is yours.
Check yourself
Question 1 / 8
0 answered
What does Claude Code's Plan mode do before you approve the plan?
Sources
- Owner
- Panaversity