Source: What’s New — Week 15 Type: Product Feature Product: Claude Code Status: Research preview Released: Week 15 (April 6–10, 2026), v2.1.92 → v2.1.101
Ultraplan kicks off plan mode in the cloud from your terminal. Claude drafts the plan in a Claude Code on the web session while your terminal stays free; you review the result in your browser, comment on individual sections, ask for revisions, and choose to execute remotely or send the plan back to your CLI for local execution.
As of v2.1.101 the first run auto-creates a default cloud environment, so there’s no Claude Code on the web setup step before you can try it.
How it works
> /ultraplan migrate the auth service from sessions to JWTsOr just include the keyword in any prompt — Claude will route it through Ultraplan.
- Terminal kicks off the plan. A Claude Code on the web session opens in the cloud and starts drafting.
- Your terminal stays free. Continue working locally while the plan generates.
- Browser review. When ready, open the cloud session in your browser. Comment on individual plan sections, ask for revisions in-line.
- Choose the execution surface.
- Execute remotely — keep the plan in the cloud, hand off to Claude Code on the web to run.
- Send to CLI — pull the plan back into your local terminal for execution.
Why it exists
Local plan mode (Shift+Tab) is great for tight loops but blocks your terminal while Claude thinks. Ultraplan separates the planning surface from the executing surface:
- Planning is async + reviewable. Plans are long-running; running them locally pins the session. Cloud-hosted means you can write code while Claude plans.
- Plans are commentable. Cloud sessions render the plan in a structured UI — point to a specific step, ask for a different approach, iterate without losing the rest of the plan.
- Execution stays where it makes sense. Some plans are best executed remotely (cloud connectors, no local file access needed); others are best brought home (touches your local repo, your environment, your secrets).
How it relates to other delegation patterns
| Pattern | Where it plans | Where it executes | When to reach for it |
|---|---|---|---|
Local plan mode (Shift+Tab) | Local terminal | Local terminal | Tight iteration; you want to see the plan inline before running |
| Ultraplan (this article) | Cloud (web session) | Cloud or local (your choice at review time) | Long-running plans you want to review async; complex multi-step migrations |
| Routines | Cloud | Cloud (durable, scheduled) | Recurring or scheduled cloud tasks |
| Managed Agents | Cloud | Cloud (durable, hosted) | Fully unattended cloud-hosted agents |
| autofix-pr | Cloud | Cloud (CI loop on a PR) | Continuous-improvement loop on an open PR |
Ultraplan is the planning-first entry to the cloud: think first, then choose where to run. Routines / Managed Agents / /autofix-pr are execution-first: define the task, let it run.
Implementation
- Tool/Service: Claude Code on the web (cloud session host) + Claude Code CLI (entry point)
- Setup: Ensure you’re authenticated with claude.ai. v2.1.101+ auto-creates a default cloud env on first run.
- Cost: Cloud Claude Code on the web session usage. Subscription tier-dependent. Audit
/costfor breakdown — Week 15 added per-model + cache-hit detail. - Integration notes: Pairs naturally with [[claude-ai/whats-new-2026-w15|
/autofix-pr]] (cloud-hosted CI loop) and Routines (durable cloud automation). Plans pulled back to CLI execute under your local environment + permissions.
Key Takeaways
- Cloud-hosted plan mode — terminal kicks it off, browser reviews it, execution surface is your choice at review time.
- Async by design. Plan generation doesn’t block your terminal; Claude Code on the web does the thinking work.
- Commentable plans. The cloud UI surfaces individual plan steps for review and revision — not a single take-it-or-leave-it block.
- Auto-bootstrap (v2.1.101+). First run auto-creates the cloud env; no separate Claude Code on the web setup.
- Decision point at review time: execute remotely (stays in cloud) or send back to CLI (run locally).
Try It
- Migration plan. Try
/ultraplan migrate <something concrete>on a real refactor you’ve been putting off. Review in browser, see whether the plan is something you’d actually accept. - Comparison vs local plan mode. Run the same prompt locally with
Shift+Taband via/ultraplan. Compare time-to-plan and plan quality. - Cloud-only execution. Choose “execute remotely” once and let the cloud run the whole change end-to-end. Useful when the plan touches infra Claude can reach in the cloud (connectors) but you’d rather not run locally.
- Plan-back-to-CLI. Generate a plan in the cloud for a tricky local refactor, pull it back to your terminal, execute against your local checkout.
Related
- Week 15 release digest — launch announcement and other Week 15 features.
- Claude Code Routines — durable cloud automation; sibling cloud-execution pattern.
- Claude Managed Agents — fully unattended cloud-hosted agents.
- CLI Reference —
--remote/--teleport/--remote-controlfor the web ↔ terminal bridges;--permission-mode planfor local plan mode. - Scheduled Tasks —
/loop(local) + Monitor tool counterparts to Ultraplan’s cloud planning. - Week 13 — PR auto-fix on web, the original cloud delegation surface.
- Opus 4.7 Best Practices — effort/thinking dynamics in long planning sessions.
Open Questions
- Cost vs local plan mode. Cloud session usage is billed differently from local. What’s the rough multiplier for the same prompt?
- Plan revision granularity. “Comment on individual sections” — how granular? Step-level, paragraph-level, plan-tree-level?
- Tier availability. Research preview — which subscription tiers can access it? Pro/Max/Team/Enterprise parity unclear from the launch post.
- Execution context when “send to CLI.” Does the local CLI execute against a fresh session or attach to the same conversation thread that initiated the plan?
- Failure handling. If cloud session errors mid-plan, is the partial plan recoverable, or do you start over?