AgentBlog

If you have used Claude Code's plan mode, you know the drill — type your prompt, wait while Claude thinks, review the plan in your terminal, then execute. It works, but for complex tasks it locks your terminal for minutes and gives you a wall of text to review in scrollback.

Ultraplan flips this. It ships your planning task to Anthropic's cloud infrastructure, drafts the plan using Opus 4.6 in a remote container, and gives you a browser-based review UI while your terminal stays free. When the plan is ready, you decide: execute in the cloud and get a PR, or teleport it back to your terminal.

How to Launch It

Three ways to start an Ultraplan session:

• Use /ultraplan followed by your prompt
• Include the word ultraplan anywhere in your message
• From a local plan's approval dialog, choose refine with Ultraplan on the web

You need a Claude Code on the web account (Pro, Max, Team, or Enterprise) and a GitHub-connected repo.

What Makes It Different

The terminal experience for reviewing plans is limited — accept, reject, or type a follow-up. Ultraplan gives you a proper document-style review surface:

• Inline comments — highlight a specific section and tell Claude exactly what to fix
• Emoji reactions — quick signal of approval or concern without writing a comment
• Outline sidebar — jump between sections instead of scrolling endlessly

You can iterate as many times as needed before approving. Claude revises the plan based on your comments, presents an updated draft, and you keep going until it is right.

The Multi-Agent Architecture

This is where it gets interesting. Under the hood, Ultraplan does not just run a single planning agent like local mode. The cloud session uses multiple parallel exploration agents plus a critique agent, all running on Opus 4.6 with up to a 30-minute compute window.

The result? Plans that are significantly more structured and thorough. And here is the surprising part — better plans lead to dramatically faster execution. In real-world tests comparing identical prompts, Ultraplan produced plans in about 5 minutes that took local mode 15+ minutes. But the bigger gap was in execution: the Ultraplan version built out a full dashboard in roughly 10 minutes versus 30+ minutes for the local plan version, because the agents had a clearer roadmap to follow.

Four Execution Paths

Once you approve the plan, you choose where and how to execute:

1. Execute in the cloud — Claude builds it remotely, you review the diff and create a PR directly from the web interface
2. Implement locally — the plan teleports back to your current terminal session with full context
3. Fresh session — start clean with just the plan as context, great for bloated sessions where you want a clean context window
4. Save to file — export the plan without executing, useful for team review or manual implementation later

The Trade-offs

Ultraplan uses more tokens up front — you are paying for cloud compute, multi-agent exploration, and richer planning. But if the plan quality means your execution phase finishes in a third of the time with fewer errors, that is a net win.

It also requires a Git repo (the cloud needs to read your codebase), and it is CLI-only — the VS Code extension and desktop app do not support it yet. Skills and custom tooling sometimes need explicit prompting to be picked up by the cloud planner.

As Abraham Lincoln put it: Give me six hours to chop down a tree and I will spend the first four sharpening the axe. Ultraplan is that axe-sharpening phase on steroids.

The Bottom Line

Planning is the highest-leverage phase of any coding task. Get it right and execution is almost mechanical. Get it wrong and you are debugging architectural decisions three hours in.

Ultraplan bets that spending more compute on planning — with better models, multi-agent exploration, and a proper review interface — pays for itself many times over in execution quality and speed. For complex, multi-file tasks, that bet pays off convincingly. For a quick bug fix, just use regular plan mode.

Ultraplan is currently in research preview and available to Pro, Max, Team, and Enterprise subscribers. Know when to reach for the big tool.