Claude Code Ultraplan: The Complete Guide to Cloud-Based Multi-Agent Planning

What Is Ultraplan?

You’re refactoring a complex system. Claude Code’s local plan mode takes 30-40 minutes, your terminal is locked, and you’re sitting idle.

Ultraplan solves exactly this problem.

It’s a feature that offloads your planning task to Anthropic’s cloud infrastructure, where Opus 4.6 gets up to 30 minutes of deep reasoning time. Your terminal is freed — you keep coding while Claude drafts the plan in the cloud.

When the plan is ready, you open it in your browser to review, comment on specific sections, request revisions, and choose where to execute it.

Status: Research Preview — requires Claude Code v2.1.91 or later.

Multi-Agent Architecture: 3 Explorers + 1 Critic

The biggest differentiator of Ultraplan isn’t just the cloud — it’s the multi-agent architecture.

How It Works

Each Explorer operates with its own independent context window — no shared state. This prevents them from converging on the same solution.

What Does the Critic Agent Do?

• Verifies completeness against original requirements
• Identifies failure modes in each approach
• Compares tradeoffs (latency vs correctness, microservices vs monolith…)
• Synthesizes the best elements from multiple approaches into the final plan

Because the Critic doesn’t participate in plan generation, it avoids self-anchoring bias — evaluating more objectively.

Speed

All 3 Explorers run in parallel on cloud infrastructure → wall-clock time roughly equals running a single sequential agent, but with significantly more diverse output.

How to Use Ultraplan

Requirements

• A Claude Code on the Web account
• Project must be a Git repository
• Claude Code v2.1.91+
• Not available on Amazon Bedrock, Google Cloud Vertex AI, or Microsoft Foundry

3 Ways to Launch

1. Slash command:

/ultraplan migrate the auth service from sessions to JWTs

2. Keyword in your prompt:

ultraplan: refactor the payment module to support multi-currency

3. From a local plan: When Claude finishes a local plan and shows the approval dialog, select “No, refine with Ultraplan on Claude Code on the web”.

Tracking Status

After launch, your terminal displays a status indicator:

Run /tasks to view session details, the link, and the Stop ultraplan option.

Review the Plan in Your Browser

When the status changes to ◆ ultraplan ready, open the session link. The review interface provides:

• Inline comments: Highlight any passage and leave a comment for Claude to address
• Emoji reactions: React quickly to signal approval or concern without writing a full comment
• Outline sidebar: Jump between sections of the plan

You can ask Claude to revise as many times as needed before choosing to execute.

Choose Where to Execute

Option 1: Execute on the Web

Select “Approve Claude’s plan and start coding” → Claude implements in the same cloud session. When finished, review the diff and create a Pull Request from the web interface.

Option 2: Send the Plan Back to Your Terminal

Select “Approve plan and teleport back to terminal” → The plan is sent to your local terminal with three options:

Ultraplan vs Local Plan Mode

Speed Benchmarks

Pros

1. Terminal stays free Keep coding, running tests, and committing — while Claude plans in the cloud.

2. Multi-agent = diverse approaches Three independent Explorers avoid single-point-of-failure in reasoning. The Critic synthesizes the best of all approaches.

3. Rich review surface Inline comments, emoji reactions, outline sidebar — far superior to terminal text review.

4. Flexible execution Choose to implement on the cloud (create a PR immediately) or send it back to your terminal.

5. 30-minute deep reasoning Opus 4.6 on cloud infrastructure gets significantly more thinking time than local inference.

Cons

1. Requires a Git repository Projects without git initialization cannot use Ultraplan.

2. Internet dependency No connection = no Ultraplan. Local plan mode still works offline.

3. Higher cost $200/month subscription. For small teams or infrequent use, local plan mode is more economical.

4. No Bedrock/Vertex/Foundry support Only runs on Anthropic’s own infrastructure.

5. Scope limitations Greenfield projects (no existing codebase) may not fully leverage Ultraplan’s analysis capabilities.

When Should You Use Ultraplan?

Use it when

• Large refactors: Restructuring modules, migrating architecture
• Clear requirements: Specs are well-defined, you need a detailed plan
• Speed matters: Tight deadlines, can’t afford 40 minutes of terminal lock
• Multi-file changes: Modifications spanning many files and modules
• Architectural decisions: Microservices vs monolith, database migrations

Skip it when

• Small changes: Fixing a typo, adding a single simple function
• Ambiguous requirements: Unclear scope → use local plan mode for faster iteration
• Budget constraints: Fewer than 5-10 complex tasks per week
• Team review needed before execution: Local plan mode offers better control
• Offline environments: No internet access

Real-World Examples

Example 1: Authentication Migration

/ultraplan migrate user authentication from session-based to JWT tokens,
including refresh token rotation, token blacklisting,
and backward compatibility for mobile clients on older versions

What Ultraplan does:

1. Three Explorers analyze the existing auth module independently
2. Each Explorer proposes a different migration strategy
3. The Critic synthesizes the final plan: schema changes, API endpoint updates, migration script, and rollback strategy

Example 2: Monolith Decomposition

/ultraplan decompose the OrderService into separate bounded contexts
for inventory, pricing, and fulfillment

Example 3: Database Migration

/ultraplan migrate from MongoDB to PostgreSQL for the analytics module,
preserve all existing queries and add proper indexing

Best Practices

1. Be as Specific as Possible

# Bad — too vague
/ultraplan improve the backend

# Good — specific with context
/ultraplan refactor the PaymentService to support Stripe and VNPay,
add retry logic for failed webhooks, and ensure idempotency

2. Use the Hybrid Approach

Start with local plan mode to define your approach → select “Refine with Ultraplan” to let Claude go deeper in the cloud. Leverage the strengths of both modes.

3. Review Thoroughly Before Executing

Ultraplan allows unlimited iterations. Don’t rush to approve — comment on specific sections, ask about edge cases, and request additional error handling.

4. Choose the Right Execution Target

• Web execution: When you want a quick PR without needing your local environment
• Terminal execution: When you need access to local services, databases, or custom tooling

5. Match the Tool to the Task

Ultraplan shines with complex, multi-file, architectural tasks. Don’t use it for simple changes — the overhead isn’t worth it.

All Claude Code Planning Modes Compared

Important Notes

• Ultraplan is in Research Preview — behavior may change
• If Remote Control is active, it disconnects when Ultraplan launches (both use the claude.ai/code interface)
• The cloud session runs in your account’s default cloud environment
• Stopping Ultraplan archives the session — nothing is saved to your terminal

Conclusion

Ultraplan represents a significant leap in how developers plan with AI. Instead of waiting for your terminal, you let 3 parallel AI agents analyze your codebase, propose approaches, and a Critic agent synthesize the optimal plan — all in the cloud.

You don’t always need Ultraplan. But for large refactors, complex migrations, or architectural decisions — it saves significant time.

Rule of thumb: If a task takes more than 20 minutes in local plan mode → consider Ultraplan.

This article is based on Anthropic’s official documentation and real-world experience with Claude Code. Ultraplan is currently in Research Preview — update to the latest version to try it out.