Written by the team developing Dure, a workspace for coding agents. Guides describe behavior documented in the linked product sources on the date shown, not hands-on benchmark results.
"Which IDE should I use for AI coding agents?" is two questions wearing one coat: which editor you type in, and where your agents run. In 2026 those have different answers, and the honest version of this guide spends most of its length on when you need nothing new at all. Everything below is documented behavior from the linked product sources, checked on the date above, and competitors ship quickly — re-read their pages before deciding.
Which question are you actually asking?
Start with where your hours went last week, not with a product name.
- You write most of the code, with help. The editor is the decision. Pick the one you are fastest in and add an agent to it.
- You delegate one task at a time and supervise it. Your agent CLI, or its own app, is the decision. Nothing else to add.
- You run several tasks at once and the queue is in your review. Isolation, sessions and review are the bottleneck. That is what an agent development environment is for.
The third case is the one people misdiagnose as "I need a better IDE."
What are the three shapes, and what does each do?
| Shape | What it is | Documented examples |
|---|---|---|
| AI-first editor | An editor where agent features are built in and you stay the main author. | Cursor calls itself "a coding agent for building ambitious software." Anthropic's VS Code extension adds "inline diffs, @-mentions, plan review, and conversation history"; its JetBrains plugin adds "interactive diff viewing and selection context sharing." |
| Terminal agent and its own app | A CLI that edits files and runs commands, now usually with a first-party desktop app. | Claude Code is "available in your terminal, IDE, desktop app, and browser." Codex runs on a CLI, an IDE extension, a cloud service, and the ChatGPT desktop, mobile and web apps. |
| Agent workspace (ADE) | A workspace around the agents: a checkout per task, sessions you can leave, a diff you can answer. | Dure is "an agent development environment (ADE) for AI coding agents." Superset runs "AI coding agents in parallel, each in its own isolated workspace." Orca gives each agent in a session its own worktree and branch. |
Sources: Cursor · Claude Code overview · Codex documentation · Dure introduction · Superset · Orca first session
These stack rather than compete. A terminal agent runs inside an editor panel; a workspace opens a worktree in whatever editor you already chose.
There are official Claude and Codex apps. Why add anything?
For many people the right answer is: don't. This section exists because most roundups skip it, and because the first-party apps changed substantially during 2026.
The Claude Code desktop app documents "parallel sessions with Git isolation, drag-and-drop pane layout, integrated terminal and file editor, side chats, computer use, Dispatch sessions from your phone, visual diff review, app previews, PR monitoring, connectors, and enterprise configuration." Each session has "its own chat history and project folder, independent of any other session," the sidebar "lets you run several in parallel," and selecting the worktree option next to the branch name gives a session "its own isolated copy of your project." You can "arrange panes for the chat, diff, browser, terminal, and file editor side by side," and SSH sessions get their own remote worktree folder. It is published for macOS, Windows and Linux. Claude Code desktop app
Codex offers worktrees in the ChatGPT desktop app, optionally running a local environment's setup scripts, with a .worktreeinclude file controlling which ignored files are copied in. They require a Git repository, start in a detached HEAD by default — so you create a branch before handing the task back — and Codex keeps the 15 most recent worktrees it manages. Worktrees "don't run locally on your phone." Codex worktrees
So if your week is one provider on one machine, the official app is a good answer and it costs nothing. Use it.
What the first-party apps are each built around is their own agent. Anthropic's documented third-party options route where the Claude model runs — Amazon Bedrock, Google Cloud's Agent Platform, Microsoft Foundry or a self-hosted gateway — rather than running a different vendor's agent CLI next to it. Third-party integrations That boundary, not quality, is what the rest of this category is answering.
Why not Orca, or another new agent workspace?
Start by admitting what is the same, because almost all of it is. Orca, Superset, Conductor, Emdash, Herdr and Dure all split work into a worktree and branch per task, run the CLI and account you already have rather than selling you a model, and end with a diff you comment on. If a page tells you a worktree per task is its differentiator, it is describing the category.
Orca is the closest comparison and a genuinely strong default: free and open source for macOS, Windows and Linux, with each agent in a session getting its own worktree and branch checked out. Orca first session
The real difference is the shape of the window, and Orca's own documentation states its model plainly: "Each worktree owns its own tab layout," and "Switching worktrees swaps the entire pane tree." Orca tabs, panes and splits That is a coherent design — one task fills the screen, and moving between tasks is a context switch. Whether you want that or the opposite is a preference about how you work, not a defect in either tool.
The other current workspaces each answer one requirement more directly than the rest:
- Scheduled jobs and programmatic control: Superset documents automations, a TypeScript SDK and an MCP server. A successful run means the workspace was created, not that the task was done correctly. Superset automations
- Shared cloud work with a team: Conductor runs cloud tasks in isolated Linux sandboxes alongside local work. Conductor cloud
- Work that starts in an issue tracker: Emdash can create a task from an issue, with a worktree and branch on by default. Emdash issues
- A terminal-native workspace: Herdr keeps sessions on a background server you detach from and return to. Herdr quick start
- Access from another device: Paseo connects mobile, desktop, web and CLI clients to the machine running your agents. Paseo overview
Our side-by-side articles go deeper on three of these: Orca alternatives, Paseo and Herdr.
What does Dure do differently, and where does it fall short?
Dure was built by people who used these tools and wanted a different window. Four documented differences, then the limits.
Different agent CLIs in one layout. The Basic launcher offers Claude Code, Codex and Kimi Code; the Beta interface mode adds Pi, Gemini CLI, OpenCode and the other registered tools. A Space can hold several of them at once — the documentation's own example layout is "two Codex agents, two Claude Code agents, Pi and a terminal." Coding agents and accounts · Spaces, panes and windows
An account chosen per agent, not per app. You add account profiles in settings and "select the intended account when starting an agent or from its account control." Rearranging your layout does not change it: "moving a pane does not switch its account." Coding agents and accounts
One layout that spans projects and hosts. A Space is "a named pane layout" you switch with ⌘1–⌘9, holding agents and terminals from different projects, local or over SSH, and you can "open a session in a larger or dedicated window." Layout is only layout: splitting a pane "does not itself create a worktree or move files," so choose a dedicated worktree when the new task needs its own files. Spaces, panes and windows · Remote projects
Returning to work after an interruption. A pane is a view; a managed session has its own host that owns the terminal and provider process. "While the session host and its machine remain running, managed work can continue when the app disconnects. Reopening Dure attaches to the current screen and later output." When an agent has ended you get its recovery actions in place, and if the task's worktree is gone, "Recreate worktree and resume" restores it first. Return to work and recover sessions
Now the limits, because they decide this for some readers. The official download is an Apple silicon Mac public beta; Windows, Linux and iOS are source-available with native validation pending, and Android is an APK preview. Built-in diff review covers local worktrees, and SSH changes go through remote Git instead. An installed CLI and provider account are required, and "features vary by provider/runtime." A reboot ends the original process, and recovery "cannot undo an already completed deployment or guarantee every provider conversation can resume." Platform status · Current limits · Return to work
If you are on Windows or Linux today, or you need a mobile app, Orca and Emdash ship those now and Dure does not.
Which requirement points to which tool?
| Your one non-negotiable requirement | Where to look first |
|---|---|
| I only use Claude | The Claude Code desktop app, on macOS, Windows or Linux |
| I only use Codex | Worktrees in the ChatGPT desktop app |
| I write most of the code myself | An AI-first editor, with an agent extension |
| I run two or more different agent CLIs side by side | Dure, or another multi-agent workspace |
| I need different accounts running at the same time | Dure's per-agent account selection |
| My code lives on a remote machine I manage | Dure or the Claude desktop app, both of which document SSH sessions |
| I'm on Windows, Linux, or a phone today | Orca or Emdash |
| I need scheduled runs or an SDK | Superset |
| I'm happiest in a terminal | Herdr, or your agent CLI alone |
A shortlist, not an exclusive capability claim and not an overall ranking.
How can you decide in one afternoon?
One repository you know, one provider, the same two tasks throughout.
- Write down which of the three questions at the top is yours. If you hesitate, you are probably in case one or two, and you are done.
- Run two independent tasks in your current setup, each isolated in its own worktree.
- Walk away for ten minutes. Time one thing on return: how long until you know which task needs you first.
- Review one diff properly — one specific comment on one line — and confirm the agent received it.
- Merge both branches and run your checks on the combined result. Count the conflicts.
Step 5 decides more than the interface does. If parallel tasks keep colliding, the fix is narrower scopes, not a different workspace. Parallel work
Questions readers ask
There's an official Claude app. Why would I use anything else?
If Claude is your only agent, you probably shouldn't. The desktop app documents parallel sessions with Git isolation, visual diff review, SSH sessions and a pane layout, on three platforms. Claude Code desktop app The case for a separate workspace starts when you want a second vendor's CLI in the same window, or several accounts running at once.
Why not just use the Codex app?
Same answer, same boundary. If your work is Codex-shaped, its worktrees in the ChatGPT desktop app cover the isolation part well. Note two documented details before relying on them: they are desktop-app only, and they start in a detached HEAD, so you create a branch before handing work back. Codex worktrees
Why use Dure instead of Orca?
Mostly for the window, not the feature list. Both give every task its own worktree and branch and use your own CLIs and accounts. Orca's documented model is one worktree's layout at a time — "switching worktrees swaps the entire pane tree" — while Dure's Space holds agents from several projects and hosts together and can open a session in a dedicated window. Orca tabs, panes and splits · Spaces, panes and windows Orca also ships Windows, Linux and mobile today, which Dure does not. Choose by which of those two sentences describes your week.
A new ADE appears every week. What actually separates them?
Four questions separate them, and none of them is "does it use worktrees," because they all do. How many different agent CLIs run in one window at the same time. Whether an account is a property of the app or of the agent. Whether one layout can span several projects and remote hosts, or one task fills the screen. What happens after a disconnect or a reboot. Ask those four about any new entrant and the list gets short quickly.
Is "lighter" or "faster" a reason to switch?
Not without numbers, and we do not publish any. Architectures do differ across these apps, but we have not measured memory or startup against another tool under matched conditions, so we make no claim. Treat any such claim — including one about us — as unsupported until you see the method.
Will a workspace make my agents better at coding?
No. The model and the CLI decide the quality of a change; the workspace decides how quickly you notice a bad one. Keep the agent and prompt identical when comparing tools, or you are measuring the model.
Does any of this remove the need to read the diff?
No, and this is the part that does not scale with parallelism. Every tool here ends with you reading a change and deciding. Dure's own guidance is to "read local diffs, send specific comments, and verify the result before combining work." Review and feedback