Mastering Claude Code Ultraplan: How to Delegate a Plan to a Remote Server and Reclaim the Terminal with `/ultraplan`

"What should I do while the AI formulates a complex plan?" This question follows whenever we start a large-scale refactoring or DB migration. While the AI agent analyzes dozens of files and formulates a plan, the terminal is tied up with the task, and the developer has no choice but to wait on the side.

In fact, this feature was revealed to the public before its official release. On March 31, 2026, an incident occurred in which 1,884 TypeScript source files were exposed due to a 59.8MB source map file being accidentally included in the npm package @anthropic-ai/claude-code v2.1.88. One of the unreleased features discovered in that leak was Claude Code Ultraplan, and Anthropic released this as the first of the leaked features to be officially released (Research Preview). It is available in Claude Code v2.1.91 or later using the /ultraplan command.

After reading this, you will be able to immediately apply the /ultraplan command to your current project and establish a workflow that allows you to perform other tasks in the terminal while planning is in progress.

Key Concepts

Differences between Ultraplan and the existing /plan

Claude Code's existing /plan command creates a plan within a local terminal session. It is difficult to perform other tasks in the same terminal until the session is complete. Ultraplan changes this structure. By separating the planning phase itself to an Anthropic remote server, it ensures that the local terminal is not tied up in planning.

Plan-Execute Separation: An architectural pattern that physically separates the planning phase of an AI agent from the phase of actually modifying the code. Agents with a separate planning phase tend to produce more stable results than those using an immediate execution method, and this is the direction Anthropic is pushing as a core differentiation strategy for Claude Code.

Ultraplan Operation Flow:

The remote server accesses the project code through GitHub integration. Since cloud containers must check out the repository to analyze file structures and dependencies, GitHub integration is a technical requirement rather than a simple account connection. For local-only repositories, using /plan is suitable.

3 Calling Methods

Method 1 — Direct Command (Recommended)

/ultraplan 모노리식 Express 앱을 NestJS 모듈 구조로 전환하는 계획을 세워줘

When you enter the command, the following response appears in the terminal, and you can then freely use the local terminal.

⠸ Ultraplan 세션 시작 중...
✓ 원격 컨테이너 연결됨
✓ 저장소 컨텍스트 로드 중...
 
계획 수립이 진행되는 동안 로컬 터미널을 자유롭게 사용할 수 있습니다.
완료되면 claude.ai에서 계획서를 확인하세요.

Method 2 — /plan Escalation

You can start locally with /plan and select cloud refinement options if you determine that the scale of the task is large.

/plan 현재 모놀리스 구조를 마이크로서비스로 분리하는 방법을 알려줘
# → 계획 수립 중 옵션 프롬프트 표시:
# [1] 로컬에서 계속 수립
# [2] Ultraplan으로 전환 (원격 정교화)

Method 3 — Keywords in standard prompt (Verification needed)

Some documentation explains that including "ultraplan" in a standard prompt automatically triggers a remote session, but this behavior may vary depending on the Claude Code version and environment. For consistent results, it is safer to use the /ultraplan direct command in Method 1.

Execution path after plan approval

Once the plan review is complete, you can choose from four paths.

Practical Application

Example 1: Establishing a Large-Scale Refactoring Plan

This involves converting an Express monolith with over 50 router files into a NestJS module structure. It is a complex task that requires dependency analysis, module separation order, and a step-by-step migration strategy.

/ultraplan 모노리식 Express 앱을 NestJS 모듈 구조로 전환하는 계획을 세워줘.
현재 src/ 디렉토리에 50개 이상의 라우터 파일이 있고,
공통 미들웨어가 전역으로 적용되어 있어.
각 라우터를 NestJS 모듈로 분리하는 순서와 마이그레이션 전략을 포함해줘.

After the plan is completed, the plan document that opens in the browser displays the following items in a structured manner.

While remote planning is in progress, you can continue working on bug fixes or tests on other branches locally. You will be notified via a browser notification when the plan is complete.

Example 2: DB Schema Migration Design

DB migration is one of the types of tasks that requires the most effort in planning because it must include a rollback strategy.

/ultraplan PostgreSQL users 테이블에 멀티 테넌시 지원을 추가하려고 해.
현재 schema.sql을 분석해서, 기존 데이터를 유지하면서
tenant_id 컬럼을 추가하는 마이그레이션 플랜을 작성해줘.
롤백 전략도 포함해줘.

backfill: The process of filling in values for newly added columns in bulk for existing records. For large tables, a separate batch processing strategy is required due to the significant performance impact.

The flow of the plan that Ultraplan establishes in this scenario is as follows:

1. Analysis of existing schema (identification of FKs, indexes, and related tables)
2. Confirmation of the Scope of Change Impact
3. Step-by-step design of migration queries
4. Establishing a backfill strategy for existing data
5. Definition of Rollback Queries and Trigger Points

If you leave specific questions as inline comments at each stage of the plan, such as "What are the criteria for the backfill batch size?", Ultraplan will suggest revisions that reinforce that section.

Example 3: API Architecture Decision Making

Ultraplan is particularly effective in architecture decision scenarios where multiple options are equally valid without a single correct answer.

/ultraplan 현재 REST API 서버를 확장해야 하는데,
REST 유지 / gRPC 전환 / GraphQL 도입 / 혼합 아키텍처 중
어떤 방향이 적합한지 트레이드오프 분석과 함께 계획을 세워줘.
현재 클라이언트는 모바일 앱과 내부 대시보드야.

gRPC: A high-performance remote procedure call framework developed by Google. It uses Protocol Buffers as a serialization format to provide lower latency and stronger type safety than REST. It is primarily suitable for internal communication between services.

For this type of request, instead of simply recommending a single option, Ultraplan presents a structured table outlining the pros and cons of each choice and the application context tailored to the current client configuration (mobile app + internal dashboard). This can be used directly as a document for team reviews or architectural decisions.

Pros and Cons Analysis

Advantages

Disadvantages and Precautions

Patterns frequently seen in practice

1. Applying Ultraplan to all tasks — For single-file modifications or simple tasks with a clear approach, local /plan or direct implementation is faster. Ultraplan proves its worth in complex tasks spanning multiple files or decision scenarios with multiple architectural options.
2. If you select Implement here directly without reviewing the plan — Even if the plan was established remotely, going through the process of reviewing it section by section in the browser and requesting corrections via inline comments helps improve the quality of the final result.
3. If you have excessive expectations regarding processing time — The time required for remote planning depends on task complexity and server conditions. As this is the Research Preview stage, it is difficult to determine a predictable range of processing time.

In Conclusion

With Ultraplan, you can run other tasks in parallel on your local terminal even while the planning phase of complex projects is underway, making a difference in both the quantity and quality of work you can complete in a day.

3 Steps to Start Right Now:

1. Version Check and Update — Check if it is v2.1.91 or higher with claude --version, and if necessary, update to pnpm add -g @anthropic-ai/claude-code@latest and check the GitHub integration status.
2. Run your first Ultraplan — Select the most complex refactoring task in your current project and run /ultraplan [작업 설명]. You can experience the benefits of the non-blocking approach by working on a separate branch while the plan is being established.
3. Utilizing Browser Review — Once the plan is complete, open the plan in claude.ai and leave inline comments on each section. You can see for yourself how this review process improves the completeness of the plan.

As this is in the Research Preview stage, we recommend testing it in a personal project first before applying it to a team project or production. If you encounter any issues or have feedback, you can report them via Anthropic Discord or Claude Code GitHub Issues.

Reference Materials

• Plan in the cloud with ultraplan | Claude Code Official Documentation
• Claude Code's Ultraplan Bridges the Gap Between Planning and Execution | DevOps.com
• Claude Code Ultraplan: Cloud Planning to Free Your Terminal | claudefa.st
• Claude Code's new Ultraplan feature moves task planning to the cloud | The Decoder
• Claude Code Ultra Plan vs Local Plan Mode | MindStudio
• How to Use Claude Code Ultra Plan: Requirements, Token Costs | MindStudio
• Everything in Claude Code's Leaked Source: KAIROS, ULTRAPLAN, Buddy and More | techsy.io
• ULTRAPLAN & KAIROS | DeepWiki
• Claude Code Ultraplan Launched: I Just Tested It | Medium
• Inside Claude Code's leaked source | The New Stack
• Claude Code Ultraplan, Why the Planning Phase Was Separated to the Cloud | Digital Focus