xcb vs Claude Code and Codex
Use Claude Code or Codex on its own when one subscription covers your work and you want every native feature. Add xcb when you pay for more than one: it runs those same tools under your accounts, sends each task to one that can take it now, and keeps the work going after you close the terminal.
Which one to pick
Pick Claude Code when
- One Claude subscription covers your work.
- You want its plugins, skills, MCP servers, subagents, web search, or IDE and desktop apps.
- You want cloud sessions, scheduled routines, or Remote Control from your phone.
- You want the agent to commit, push, or run native macOS builds.
Pick xcb when
- You pay for more than one of Claude, Codex, and Devin, or hold more than one account, and want one place to send work.
- You want each task sent to an idle account that isn’t at a known limit, and held there until the run ends.
- You want tasks in several projects to keep running after you close the terminal, with one place to answer their questions.
- You want another agent to hand off work with
xcb --json route.
Side by side
| Aspect | Claude Code | xcb |
|---|---|---|
| Model access | A Claude plan, an API key, or a cloud provider such as Amazon Bedrock | Your subscriptions, through each provider’s own tool; xcb has no model access of its own |
| Accounts | One account per sign-in; switch with /login | Several accounts across providers; each task goes to one that can take it now |
| Tools in a run | Built-in file, shell, and web tools, plus MCP servers, plugins, skills, and subagents | Claude Code’s built-in tools are off; xcb’s file tools and command VM take their place |
| Sandboxing | An optional OS sandbox for shell commands: Seatbelt on macOS, bubblewrap on Linux and WSL2 | The whole Claude Code process runs in a sandbox with a private home directory |
| Commands and Git | Runs commands on your machine and can commit and push | Commands run in an offline Linux VM on macOS ARM64; Git is read-only for workers |
| After you close it | A local session ends; cloud sessions and routines run in Anthropic’s cloud | Managed tasks keep running under a background supervisor on your machine |
| For other programs | claude -p and the Agent SDK | xcb --json route takes one task and returns the result, the account and model used, and how the run ended |
| Platforms | macOS, Linux, WSL, and Windows | Claude runs on macOS ARM64; Claude on Linux x86_64 isn’t confirmed, and there is no Windows build |
Codex
The same choice applies to Codex, OpenAI’s open-source coding agent for the terminal, IDE, and desktop, which signs in with a ChatGPT plan or an API key. Codex has its own sandbox and approval modes, subagents, skills, MCP servers, and cloud tasks. Use it on its own when one ChatGPT plan covers your work.
xcb runs Codex through its app-server ↗, which OpenAI documents for building Codex into your own product, with ChatGPT device sign-in. Inside an xcb run, Codex’s own shell, web search, and subagents are off, and xcb’s file tools take their place.
xcb runs Codex on macOS only and supports specific Codex builds, adding new ones after they pass its checks, so it can trail the newest Codex release. A coding session on a signed-in Codex account passed on the previous supported build and hasn’t been rerun on the current one.
Your Claude plan through xcb
xcb runs the unmodified Claude Code binary, signed in through Claude Code’s own claude setup-token flow. It keeps the resulting token in its private state on your machine and passes it only to Claude Code. Anthropic’s Claude Code legal page ↗ says its credential rules don’t “prevent an end user from signing in to the unmodified Claude Code binary with their own Claude subscription”. The same page says Pro and Max usage limits “assume ordinary, individual usage of Claude Code and the Agent SDK.”
xcb runs Claude Code in print mode, and Anthropic’s help center ↗ says claude -p usage still draws from your plan’s usage limits. Anthropic paused a planned change to that and says it will announce any update before it takes effect. Each task uses the limits of the account it runs on; xcb doesn’t raise or pool them.
Sources
- Claude Code overview ↗Checked
- Claude Code sandboxed Bash tool ↗Checked
- Claude Code routines ↗Checked
- Claude Code Remote Control ↗Checked
- Claude Code authentication and `claude setup-token` ↗Checked
- Anthropic: Claude Code legal and compliance ↗Checked
- Claude help center: the Agent SDK and `claude -p` on your plan ↗Checked
- Codex on GitHub ↗Checked
- Codex app-server ↗Checked
- xcb: Accounts and modelsChecked
- xcb: Tests, builds, and recoveryChecked