Use a terminal agent for scripted and CI work, an IDE agent when you want inline diffs while you edit, a desktop app to run sessions side by side, and a web or cloud session for long tasks you can leave. The agent engine can be the same across all of them. What changes is control and where code runs.
The surfaces
Anthropic's Claude Code overview shows the pattern. It runs in the terminal, in VS Code and JetBrains, in a desktop app and on the web, and says each surface connects to the same underlying engine, so repo instructions, settings and MCP servers work across all of them. OpenAI's Codex CLI is described as a terminal tool that can also run non-interactively in scripts and CI. GitHub Copilot's agent mode works inside editors such as VS Code, Visual Studio, JetBrains, Xcode and Eclipse.
Terminal. Best for automation. The Claude Code docs show piping logs into it and running it in CI, and Codex documents a non-interactive exec mode. It is also the least distracting place if you live in a shell.
Trade-off: diffs are text. You need to be comfortable reading them and approving commands.
IDE. Best when you want to see changes in context. The VS Code extension is described as offering inline diffs, mentions, plan review and history in the editor. You approve edits where you can already see the surrounding code.
Trade-off: the agent shares the editor's trust model. VS Code's workspace trust and MCP trust dialogs matter here.
Desktop app. Best for parallel work. Anthropic's overview says its desktop app lets you review diffs visually, run several sessions side by side and schedule recurring tasks.
Web or cloud. Best for long, self-contained tasks. You start a job and check back later, even from a phone. Anthropic's security page says cloud sessions run in isolated VMs with limited network access by default. Its remote control feature is different: the web interface connects to a process on your own machine, so execution and files stay local.
How to choose
- Run in CI or a script. Surface: Terminal.
- Review small edits live. Surface: IDE.
- Several tasks at once. Surface: Desktop app or cloud.
- Long task while you are away. Surface: Web or cloud session.
- Strong isolation from your laptop. Surface: Cloud VM or a local sandbox.
Safety is separate from surface
Whichever you pick, set permissions, restrict MCP servers and review diffs. See AI coding agent governance. Safeguard has CLI, VS Code, JetBrains and Open VSX surfaces for scanning and policy checks, listed on its download page.
Quick check
Do my MCP servers carry across surfaces?
For Claude Code, Anthropic says repo instructions, settings and MCP servers work across its surfaces. Other tools differ, so check each one.
Is a cloud session safer than local?
It is isolated from your machine, but you must trust the environment and what it can reach. Local work needs its own sandboxing.
Sources
- Claude Code overview
- Claude Code security
- Codex CLI
- GitHub Copilot MCP setup
- Safeguard download page
Last reviewed: 19 October 2026.
