Anthropic shipped a research preview of Ultraplan, a feature that hands Claude Code's planning phase off to a cloud session while your terminal stays free. The idea: reading a codebase and building an implementation strategy doesn't need your local machine. Writing the actual code often does. The official docs went live alongside CLI version 2.1.91, which is the minimum required.
Three ways to trigger it. Run /ultraplan with a prompt. Drop the word "ultraplan" into any normal message. Or start a local plan, then pick "Refine with Ultraplan" from the approval dialog, which ships your existing draft to the cloud as a starting point. That last path skips the confirmation step. While the cloud session runs on Anthropic's Cloud Container Runtime with Opus 4.6, the terminal shows a status indicator: drafting, needs clarification, or ready. Anthropic engineer Thariq confirmed on X that token consumption is roughly comparable to standard plan mode.
The browser review surface is the real selling point. Developers can leave inline comments on specific plan sections, react with emoji, and navigate via an outline sidebar. Each revision generates a new draft, and the loop runs as many times as needed. Once approved, two paths: "Start coding" keeps everything in the cloud session and produces a PR through the web UI, or "Teleport back to terminal" archives the cloud session and drops the plan into your CLI with options to inject it into the current conversation, start fresh, or save to a file.
Requirements are specific: a Claude Code on the web account, a GitHub repository, and CLI v2.1.91 or later. It won't work over Amazon Bedrock, Google Cloud Vertex AI, or Microsoft Foundry. Active Remote Control sessions disconnect when Ultraplan launches, since both features share the same interface.
Bottom Line
Ultraplan requires a GitHub repo and Claude Code on the web account, and runs as a research preview with token costs comparable to local plan mode.
Quick Facts
- Minimum CLI version: 2.1.91
- Runs on Opus 4.6 in Anthropic's Cloud Container Runtime
- Up to 30 minutes of dedicated cloud compute per session
- Not available on Bedrock, Vertex AI, or Microsoft Foundry
- Token consumption roughly matches standard plan mode (company-reported)
