If you spend all day with agents, the interface you run them through stops being a detail and becomes the whole experience. I've been bouncing between every option Anthropic offers — the CLI, the VS Code extension, the desktop/web app — plus a few third-party wrappers. Each one is great at something and frustrating at something else. Here's how I think about the trade-offs, and where I've landed.
The terminal
The CLI is the canonical way to use Claude Code, and for a single agent doing a single thing it's hard to beat. One window, one task, you watch it work.
The problem shows up the moment I want to run several agents in parallel. Then the terminal starts to fight me. I need multiple tabs, the terminal pinned somewhere, VS Code open on the side just to see which files changed. If I want a second conversation I open another terminal — or lean on something like Ghostty or tmux for tabs, which works but adds friction. Organizing several agents at once is awkward, and so is browsing history: I'm typing commands all the time when a single click would do. Sometimes I'd rather click an entry in a history panel than type /resume. Personal taste, sure — but at scale the CLI isn't where I'm most comfortable.
The VS Code extension
The extension is excellent for quick, iterative work — small changes, fast loops, everything documented right there. And I like having the code physically close: I can see which files the agent is touching, manage my Git branches, make commits, keep a terminal within reach.
The catch is that it lives in the sidebar. It's never quite the main event. Honestly, I barely touch the actual code in VS Code anymore — at most I'll edit a markdown file or tweak a prompt. So I'd love the agent to be more central and the editor to be the supporting cast, not the other way around.
It also trails the console on features: no mobile notification when a run finishes, the newest models tend to land there last, the odd command is missing, and multi-agent management isn't great — you're stuck hitting the history button to see everything instead of having all your agents in view at once.
The desktop / web app
On paper this is the ideal. All your agents in one place, easier to manage them down the left side, you can open a terminal, open files, and it ships with the latest models and the best commands. Best of all, it pings my phone the moment it needs a permission, so I can step away and keep working from mobile. That part is genuinely close to perfect.
But it has gaps. I can't edit files directly — if I want to add one small line to a CLAUDE.md by hand, I can't; I have to ask the agent, or keep VS Code installed and open anyway. And then there's branching: it offers automatic PRs and Git worktrees, which sounds great, but I'm not used to it yet. With commit and push I can see exactly what's happening; here it feels a bit more opaque and uncomfortable. I'm fairly sure that's a me problem rather than a flaw — this is clearly the intended path, it spins up worktrees effortlessly, and it's probably the best way both to ship and to actually learn that workflow.
Third-party tools
I'm not talking about Cursor, Windsurf or Antigravity here — those are separate subscriptions, and if you want to use your full quota it makes more sense to pay for Claude directly. I mean wrappers like Conductor that make parallel agent work easier. I like what they do, but they make me uneasy: they're not the first-party provider. They won't have the latest updates the second they ship, they can carry more bugs, and they could start charging extra for things down the line. Yes, they run Claude Code under the hood — but I'm not fully comfortable depending on them.
Where I landed
So my current split looks like this:
- VS Code for small edits and the little things.
- Claude Desktop for big refactors and large new features — when I want all my agents in one place and notifications on my phone.
- The CLI when I'm SSH'd into a machine and want to leave a long process running: a heavy refactor, a big experiment, something I can fire off and not babysit. That's usually the more exploratory lane — chasing fast results rather than polished production code.
None of them is the one true answer. The honest takeaway is that the "best" interface depends entirely on what kind of work you're doing — quick iteration, serious building, or unsupervised experiments — and the trick is knowing which one to reach for.