Desktop apps for coding agents compared: Caprock, Orca, Conductor, Claude, Codex, Warp and Nimbalyst

Seven apps now run Claude Code, Codex and other coding agents side by side. Caprock, Orca, Warp and Nimbalyst are on all three desktop systems; Conductor is on macOS; Claude's and ChatGPT's desktop apps each run their own vendor's agent. They differ most in licence, in how a phone reaches your machine, and in what they show of cost and plan limits.

Claude Desktop alternatives for coding

Claude's own desktop app runs Claude Code in its Code tab, on a paid Claude plan, on macOS and Windows with Linux in beta. To run other agents beside Claude Code in one window, three picks from the table below, read on 7 October 2026:

1. Orca. MIT, on macOS, Windows and Linux. Runs Claude Code, Codex, OpenCode, Gemini and more, with usage and plan limits for eleven providers.
2. Conductor. macOS only. Runs Claude Code, Codex, Cursor and OpenCode; free locally; Pro $50 a person and Teams $60 a user a month, read 27 October 2026.
3. Caprock. Apache-2.0, on macOS, Windows and Linux. Claude Code, Codex, OpenCode, Gemini CLI, Grok Build and Hermes Agent in tabs per project; the sessions it did not start recorded too; every pull request followed from agent to release with its agent cost; a phone from anywhere through an end-to-end encrypted relay, with no account. Free. Download Caprock

Side by side

Cost, limits and price

Notifications and a system prompt per project

Every one of these apps works with git worktrees, one per session (Warp leaves creating them to you), and all but the ChatGPT app and Nimbalyst document two sessions or terminals side by side (Conductor's split tabs are an experimental setting). Two things differ more, read 7 October 2026:

• Answering a permission prompt from the desktop notification. Caprock does it on macOS and Linux, for Claude Code sessions it started. Orca, Warp and Nimbalyst notifications open the session; Claude desktop notifies when a task finishes; the ChatGPT app's permission notifications can be switched on or off, and approving from afar is done in its phone app. Conductor documents no notification approvals.
• A system prompt per project, set in the app. Caprock calls it Project instructions and appends it to the system prompt of every Claude Code session it starts in the project, and to the developer instructions of every Codex session it starts there, and passes it to Grok Build as its rules; OpenCode, Gemini CLI and Hermes Agent sessions do not get it. Conductor has repository-wide "General preferences" for its agents. Claude's Projects carry project instructions, in a public beta on Pro and Max. The ChatGPT app's projects carry instructions for its chats, while its local Codex projects point to AGENTS.md or checked-in documentation. Orca, Warp and Nimbalyst document only the agents' own files.

Download and install size

Measured on 11 October 2026 from the files each project publishes on GitHub: Caprock v0.100.0 and Orca v1.4.224. Orca is an Electron app and ships its own Chromium; Caprock is a Go daemon in a Tauri shell, which uses the web view the system already has. The Linux AppImage is the exception: it carries its own WebKit, so the gap there is small.

Caprock's macOS file is one universal build for Apple and Intel chips; Orca ships one per chip. Installed on the same Mac, Caprock.app takes 47 MB and Orca.app (1.4.220) 624 MB, by du. These are sizes on disk, not memory or speed; those depend on how many sessions run and are not measured here.

What each one is

• Orca — an open-source app built around worktrees and the agents' own CLIs, with iOS and Android apps. A phone pairs through Orca Relay, which needs a sign-in, or directly over a LAN or Tailscale; mobile is marked beta.
• Conductor — a Mac app that runs Claude Code, Codex, Cursor and OpenCode in parallel workspaces. It needs GitHub; its iOS app, released 2 October 2026, works with cloud workspaces, and pricing lists it under Pro. Since 0.91.2 it can also control local workspaces on your Mac, an experimental setting.
• Claude desktop, Code tab — Anthropic's app for Claude Code, on a paid plan (Pro, Max, Team or Enterprise). Remote Control lets the Claude phone app drive a session, through the Anthropic API.
• The ChatGPT desktop app — Codex is one of its two workspaces; Remote reaches a Mac or Windows PC from the ChatGPT phone app through a secure relay. It is what OpenAI's Codex app became: Codex desktop app.
• Warp — a terminal that runs third-party CLI agents beside its own; Remote Control shares a running session to a link you open on a phone.
• Nimbalyst — an open-source workspace with visual editors around Claude Code, Codex and Gemini, with iOS and Android apps that sync end-to-end encrypted, with an account.

Claude Squad, Vibe Kanban and Superset

Three names come up in the same searches. None is a pick here; each was read on its own pages on 27 October 2026.

• Claude Squad (AGPL-3.0), a terminal app, cs, that runs Claude Code, Codex, Gemini and other local agents each in its own tmux session and git worktree, is unmaintained: its last commit and its last release, v1.0.20, are from 20 August 2026, and a maintainer wrote in March that both authors work full-time jobs elsewhere. The Windows build does not run — "error starting tmux session: unsupported", an open issue.
• Vibe Kanban (Apache-2.0), a kanban board in the browser with a workspace per agent, shut down on 10 April 2026: bloop wrote that "the vast majority are free users and we couldn't find a business model", refunded its Pro plan and removed the hosted issues, comments and projects. One former employee maintains the repository and local workspaces still run.
• Superset became Paradev on 26 October 2026: Pro $20, Team $60 and Ultra $200 a user a month, on macOS, with Linux and Windows open requests. Its status view covers only agents started from its own terminals ("The hooks do nothing in other terminals", its usage doc).

Where Caprock differs

⤢

⤢

Every app here is a window around the agents it started. Caprock's difference is the record under the window: a background service that records every session on the machine, the ones you start in your own terminal included, with limits across subscriptions, each pull request's CI state and what it cost.

• Cost for every agent it reads. Claude Code, Codex, OpenCode, Gemini CLI, DeepSeek Harness and Hermes Agent land in one table, priced per repository; Claude and Codex plan limits sit in the menu bar, on the Home screen and on the Cost screen.
• Sessions from your own terminal count. A claude started outside Caprock appears with its cost; Caprock types only into sessions it started.
• A phone with no Caprock account. It pairs by QR code and reaches the computer from anywhere, through an end-to-end encrypted relay. A terminal picks up from the last byte it had, and the chat view orders messages as the session produced them, however late they arrive.
• Several tries at one task, by cost. One agent gets the same first message in two to four copies of the repository, and Compare sets them side by side: files and lines changed, what each has cost so far, how long it has run.
• Approve or Deny from the notification. On macOS and Linux, a permission prompt from a Claude Code session Caprock started arrives with Deny, and with Approve only when the notification shows the whole request.
• Measured, not claimed: 0.44–0.59 s from launch to a ready window with the background service running, five runs on one M1 Pro Mac, 2026-10-05.

What Caprock does not do

• No code editor. It runs the agents' terminals and shells; you edit in your own editor.
• No cloud workspaces, no SSH remotes. Sessions run on your machine, or on your other computers that run Caprock, linked through its relay.
• Buttons in notifications are macOS and Linux only. On Windows you answer on the session's prompt card.

Which to use

• Caprock — one record of every agent in every repository: the sessions it did not start, limits across subscriptions, cost per pull request, history; and a phone with no account. Install it.
• Orca — the same idea, with native phone apps: Orca vs Caprock, row by row.
• Conductor — a Mac, GitHub-centred workflow, with cloud workspaces; pricing lists those and the mobile app under Pro.
• Claude desktop or the ChatGPT app — you use one vendor's agent and want its own interface.
• Warp — you want a terminal first, with agents inside it.
• Nimbalyst — you want documents, mockups and diagrams beside the agents.
• cmux or Herdr — you want agents in a terminal rather than an app: Herdr vs cmux, and cmux on Windows if you are not on a Mac.
• Claude Squad or Vibe Kanban — neither: one is unmaintained and the other's company shut down, above.

A GUI for Claude Code alone, official or open source: Claude Code GUI. Choosing the agent rather than the app: Claude Code compared with Codex, Cursor, OpenCode, Gemini CLI, GitHub Copilot, Antigravity and Kiro.