I spent years in Vim, Neovim and keyboard workflows that never touch the mouse. Had you told me I'd end up using Claude Code Desktop, an Electron-based desktop app, every day, I wouldn't have believed you. That's where I am, and I don't regret it.
TL;DR Import your terminal sessions into Claude Code Desktop from Help > Troubleshooting and try it for a week or two. Your important conversations come along with their context.
In June I was still running my agents from the terminal, with Warp and Neovim, and I had already written about five features worth leaving the terminal for. This article is what came next: what I use in Desktop every day, when I reach for it, and how my cloud scheduled tasks work while I sleep and leave me pull requests by morning.
It's for two kinds of reader: you come from the terminal and going back to a desktop app feels like a regression, close to heresy; or you're new to Claude Code and can't tell whether to pick the terminal or Desktop.
My read, and it is an opinion: Anthropic is putting more care into the desktop app than into the terminal. I write about Claude every day and follow the changelog closely, and the attention shows more in Desktop. That doesn't mean they're dropping the terminal.
Import your terminal sessions first
In Claude Code Desktop, open Help, go to Troubleshooting and look for the option to import CLI sessions. It brings in every session you have in the terminal.
For people like me it's extremely important. Rather than starting lots of new sessions, I keep a few very important ones, with aggressive compaction and context I want to hold on to. I had done it by hand, asking Claude, before I knew Help had its own button for it.
Chat, Code, and all of Claude in one place
At the top you have two tabs, Chat and Code. Both are conversations. Code ones are usually tied to code, to repositories, though you don't need a repository to use the tab. The way I see it, Chat is the conversational side: trivial things, web research, no projects or code orchestration behind it. Code is for producing, validating and deploying code.
I think the deeper reason Anthropic cares about Desktop is the product itself, which has been somewhat scattered: Claude Design, Claude, Claude Cowork and Claude Code in the terminal, each on its own. Desktop tries to put it all in one place. It's not for everyone, and you can keep using Claude Code separately, but putting it together creates synergies you'll see in the rest of this article. Claude Design, for one, leaves its separate site on December 14, 2026 and moves into Claude.
Environments: local and cloud
The mental model that ties all of this together is the environment, and Desktop puts it on the home screen. Other harnesses, like Codex or Cursor, use something very similar.
An environment is a set of folders, one or several, and those folders may or may not hold repositories. The usual case in the Code tab is one or more GitHub repositories, but it isn't required. My two main environments are wmedia, the site where I write about Claude Code every day, and AI Infra, the repository with the plugins and skills I use, most of them for work.
The gear on each environment shows its folder and, when there's a repository, its branch, and lets you add more folders. Some environments have dependencies spread across several folders, and each one may hold its own repository, or not, and even work in worktrees. You decide the layout case by case.
There are local environments and cloud environments, and I see Anthropic putting a lot of effort into getting you to try the second kind. A cloud environment recreates a local one: its secrets, its env files and the spec of what it needs to run with its dependencies. A clone of your local environment, in the cloud. So I can start a task in the cloud, close the laptop and get on with my day. Or start locally and continue in the cloud: right-click the session in the list, Open in, and continue in the cloud, as long as that environment exists. If it doesn't, Claude helps you recreate it. Each destination is covered in Claude Code Desktop environments, and what a session takes along in moving a session to the cloud.
The session panels: environment, context, changes and connectors
Cmd+B (Ctrl+B on Windows and Linux) shows the sidebar with your conversations. When you start a new one you pick the environment, with or without a worktree, or any folder on your file system, repository or not. If there's work in progress in that environment, Claude will probably infer you want to continue it; in my case it asked. And not everything needs Opus on high: for that conversation I dropped it to medium.
On the right you get panels with information, which you can reorder by dragging:
- Environment: name, folder and active branch. You can ask Claude to add another folder, or move the session with
cd. It works, but it has side effects and, in my opinion, is still a bit confusing: it shows where you started and where you are now, and at some point you lose track of where you are. - Remote control: you share the conversation and carry on from the iOS or Android app as if you were at the desk. Your machine has to stay on, unless it's a cloud task. I use it a lot at the gym. More in controlling Claude Code from your phone.
- Context: messages, MCPs, system prompt and skills. In the terminal and in Desktop alike,
/contextgives you roughly the same. I mostly watch the messages, which are what you clear with/clearor compact with/compact. In the session from the video it sat at 28% of 400,000 tokens, because my autocompact is set to 400,000. That's my sweet spot: past it, context rot starts. Some people use 200,000, which feels a bit aggressive, even to me. How that window gets decided is in Claude Code auto-compact. - Changes: what was added and removed (3 and 1 in the video) and the button to create a PR.
- Connectors: plugins and MCPs, including the ones you had connected at some point. You can also reach them from the + icon on the left, but here they're at hand and change with the conversation and the environment.
Inside the conversation I use two small things all the time. The first is pinning a message: in a long conversation, with several messages from you and from the orchestrator, you pin the one you want to come back to and it shows up in a table of contents. Each message's menu also has copy, fork and read aloud.
The second is asking about a part. You select a piece of a response, or of several, and ask about that piece: clarify this, explain that, change this other thing. As far as I know, you can't do that in the terminal. I use it on long responses, to reply with something more articulate than yes or no.
Two shortcuts: Cmd+K and Ctrl+Tab
These are the only two you need to memorize, in my view.
Cmd+K (on Mac) opens a contextual fuzzy search, like any modern editor or IDE: it finds sessions, artifacts, projects and scheduled tasks. Start typing and it finds it. As far as I know, the terminal doesn't ship this; with a multiplexer you can build something even better, but you build it yourself.
Ctrl+Tab jumps between the two conversations you're moving between. Hold it and you keep going down the list. It's a common UX pattern in developer tools, and here it comes built in.
I don't usually recommend working in two sessions at once, because I try to focus on one project, though the more AI assistance you use, the more you drift into multitasking. If you want to, Option + click on a conversation opens it as a second column, and another as a third; you can arrange them in a grid. Cmd + click opens it in a separate window you can pin. I don't use it much, though sometimes I've had one floating there to keep an eye on what's happening without giving it my full attention. Once you have several, Claude tells you which did what.
Whiteboard: agree on the architecture before Claude builds it
Whiteboard is recent and I believe it's still in beta. I used to fake it with a plugin of my own; now it's native and I use it almost every day.
In the conversation for the right environment I asked it to draw on the whiteboard, simple and stripped down, how the site's opt-ins work today, across several lead magnets and several entry points. The canvas opened with a diagram of that piece of the architecture. Any panel on the right can be expanded, and on the whiteboard you pan with the hand tool, as in FigJam or Excalidraw.
What matters is that it's conversational. You can look, edit and add, but mostly I ask. I asked whether we had logs for that flow; it went to check and told me we didn't. Move something and Claude sees it and reacts. It gives you a shared understanding of an API surface or a part of the architecture: you make sure you're both talking about the same thing before Claude builds it, and the sooner you go through this, the better. I'm a very visual learner, and having the visual part be conversational helps me a lot. More in Whiteboards from Claude Code.
Alongside Whiteboard I use Visualize a lot. I asked for the diagram of the reCAPTCHA flow on my site and it drew it right in the conversation, without leaving the app. You can download it as SVG or PNG, copy it or ask for changes, and if you open it in the browser you can draw on it and Claude sees it. I know the terminal can show images under very specific conditions, so I won't call it impossible, but I don't think it's this fast or this practical.
Side conversations: conversations that report to the main one
I use them constantly. I ask Claude to create a side conversation for a task. It's a nested conversation with a specific prompt; when something significant changes, it reports to the main session, and that session reports to me. It's half delegation: I can step into it whenever I want, but it has a narrow goal. Sometimes it's a topic rather than a goal: my RSS, YouTube, routines and funnel side conversations started with one thing and stayed because they cover a topic. All of them report to Plan, my main conversation, in one way or another, so Plan knows everything.
It isn't autopilot. That would be /subtask, a kind of fork for a subagent that works fully on its own (the subagent that inherits your context), and subagents are run by the orchestrator, not by me. It isn't a fork either: a fork clones the conversation with all of its context (/fork in Claude Code). I don't use it much, because side conversations already report to the main one and what matters always arrives.
A real example, from the conversation I recorded in the video. Search Console suggested the opt-in flow might be failing on mobile. I asked Claude for a side conversation that would use Claude in Chrome to check the 51 tips opt-in on the home page on mobile and leave screenshots on my desktop. Claude in Chrome or the DevTools MCP, either was fine by me (how the two differ). The side conversation showed up in the list with my request as its prompt, did the job, and the conversation I launched it from received its message. For real work I would have gone further: a test subscriber, the whole flow in English and in Spanish, the logs.
Messages also go top down. I asked Claude to ask that side conversation for its verdict, and it replied that it looked fine. And they work between any two conversations, not only within one hierarchy: from that same conversation I asked the León AI conversation, about the meetup I run in León, when the first session is and whether everything was more or less ready. In this case it didn't make much sense, and it's not for every day, but conversations stop being silos. At some point you may want to archive or delete a side conversation: what matters stays in the conversation it hangs from.
One requirement: every conversation has to be in auto mode or bypass permissions; otherwise the message waits in a queue until the conversation is free. I work in YOLO mode almost all the time, so it doesn't affect me. If you'd rather not go that far, look at auto mode.
Terminal, diff, browser and files: everything goes in as context
The panels on the right aren't inert: they're connected to the session, and all of them let you attach what you see as context for the conversation.
- Terminal: good enough, though not the best, and it has no Vim mode. I'm using an app without Vim mode, which is unbelievable coming from me. I don't use it much, but it's connected: Claude can type into it and sees what commands return. Run a
curland ask it to read the terminal output, or copy the output and attach it as context. - Diff: what changed on that environment's branch or worktree. You can add a comment or attach it as context. I attach: I pick pieces, ask why this, explain that, and Claude gets all the questions at once and answers.
- Browser: not the best either, but it has an MCP built in and it's connected to the session. When the deploy is simple, I render it right there. Next to the other options, the Desktop browser is not your Chrome.
- Files: not the best viewer and no IntelliSense, but for a quick edit without opening your editor it's enough. For me it covers 70 to 75% of edits.
That leaves the side chat, which is basically /btw with its own panel and which I don't use much, an iOS simulator I don't use, and artifacts, with their list. All of Claude is in here.
Cloud scheduled tasks: pull requests while I sleep
My scheduled tasks marked as cloud run even when my computer is off. Each one is generally a cloud task with a cloud environment. It lives on Anthropic's servers: it brings the environment up, clones the repository, runs the task and returns the result. The clone lasts up to 7 days, I believe; if it's gone, it gets cloned and recreated.
Each task shows its run history and, under details, the instructions, which are the prompt. It wakes up in different ways: on a schedule, on a GitHub event or through the API, and mine are scheduled. You add the connectors it needs, notifications (I get push; some people use Slack), the model and reasoning effort, the cloud environment and the repositories.
Every run starts as if it were the first: it doesn't remember earlier ones. This isn't Managed Agents on the Claude platform, and you're the one who gives it memory. I use a Notion document the tasks update with a bullet point (how to give them a notes page). I have cloud environments for FrontendLeap, Platform and wmedia, clones of what I had locally.
Many mornings I wake up to push notifications from Claude with PRs made on my behalf. One of them added the banner at the bottom of my tips that suggests adding the site as a preferred source on Google. Whatever you hand to this system has to be autonomous, so I often start with CRAFT's shaping: I work out what I want, and I delegate once a human is no longer needed and there isn't much uncertainty left. The task opens the PR with a screenshot. Since I have the Claude app installed, I comment ("what other copy options did you consider?") and the task wakes up again and answers; if I ask for changes, it makes them. That PR belongs to a row in Notion, which is what I use. The full mechanism is in scheduled tasks in Claude Code and in keep talking to a finished run.
I'm the bottleneck now: my capacity to validate sets the ceiling on what I get done, not the code I write.
My verdict
Claude Code Desktop isn't for everyone. If your terminal setup took you years to build, I get the reluctance. But if you want an all-in-one that's good enough and works quite well, it may be an option worth trying. I run Claude personal and Claude Work at the same time, sometimes with conversations open in both, and it holds up. It's Electron, yes, but it isn't excessive.
Import your conversations, try it for a week or two, and tell me how it goes.