Claude Pro/Max

Claude Pro/Max is an optional personal model connection. A person connects their own Claude plan, and a thread they pin to Claude Code runs with Claude Code as its agent in the thread’s Orb. Model use bills to their Claude plan. The deployment needs no configuration and pays no model cost for these threads.

Connect a plan

On your own computer, install Claude Code and run claude setup-token.
Sign in with your Claude Pro or Max account and approve access. The command prints a token that starts with sk-ant-oat. It is valid for one year.
In dx, open Personal Settings → Model Routing and choose Add → Claude Pro/Max.
Paste the token and choose Connect Subscription.

dx checks the token with Anthropic before storing it and refuses a token Anthropic rejects. It encrypts the token like an API key and never returns it. Each person has one plan connection; adding it again replaces the token. Workspace routing does not offer it.

Start a Claude Code thread

In the new-thread mode picker, choose Claude Code, either the lamp under the dial or the Claude Code row, then a model and an effort. Claude Code models are not listed among raw models or Mode Dial slots, and custom connections cannot serve them.

Claude Code works with its own tools, such as Bash, Read, Write, Edit, web search, and subagents. dx adds its own tools as mcp__dx__*: the thread tools, including find_thread and read_thread, Code’s code_exec and tool_search for connected MCP servers, skill resources, and installed plugin tools. The transcript, Stop, Retry, Changes, and Files work as in any other thread.

Plan limits and extra usage

Claude plans have usage limits: a 5-hour window, a weekly window, and on some plans weekly limits for one model (the Opus, Sonnet, and Fable limits). A Claude account can also have extra usage turned on, which Anthropic bills at API rates once a limit is used up.

The Claude Pro/Max card in Model Routing shows the limits Anthropic reports for your plan, with when each resets. dx reads them from Anthropic when Model Routing loads and when you choose the refresh button beside Plan Usage. While a limit is used up and extra usage could be billed, dx shows the last reported values instead.

The card’s Claude Usage choice decides what happens at a limit:

Subscription Only (the default): dx stops Claude Code when Anthropic reports that extra usage is paying, and does not start it again until the limit resets. A weekly limit for one model, such as the Fable limit, holds only threads on that model. The response in flight when Anthropic reports it may already be billed as extra usage. To make sure nothing is billed, turn off extra usage in your Claude account settings.
Allow Extra Usage: Claude Code keeps working on extra usage.

A thread that reaches a limit shows a Claude subscription limit reached card above the composer with the reset time. Choose Retry after the limit resets or after you allow extra usage.

Orb requirements

The standard Orb image includes Claude Code, Python 3 for dx’s tool server, and awk. A workspace whose Claude Code is missing or older than 2.1.310 installs the current release the first time it runs a Claude Code thread. A custom image needs network access to claude.ai for that install, or Claude Code on the PATH.

Token handling

The token reaches Claude Code only on a file descriptor for one run, and only for the thread’s pinned Claude Code model. dx removes ANTHROPIC_* and CLAUDE_CODE_USE_* variables from Claude Code’s environment and passes command-line settings that fix the API URL and clear proxy and CA variables. Workspace variables and a repository’s .claude settings therefore cannot point Claude Code at another API URL or proxy. Hidden secret values Claude prints are redacted from the transcript. Commands that a repository’s setup or the agent runs in the Orb run as the Thread’s owner, so people should connect a plan only for repositories they trust.

Troubleshooting

“Anthropic rejected this token”: the token was mistyped, revoked, or expired. Run claude setup-token again and replace the token.
“Anthropic could not be reached”: retry later; dx stores no token it has not checked.
The Claude Code row shows Connect Claude Pro/Max: connect a plan first, in personal Model Routing.