DEEP DIVE · 01
Claude Code Desktop
the full tour
The Code tab inside Claude Desktop, redesigned 2026-04-14 (GA). Parallel sessions, Plan mode, MCP, visual diff. Mac and Windows, no terminal needed.
GA
Mac
Windows
Pro / Max / Team / Enterprise
1. What Claude Code Desktop actually is
Claude Code Desktop is the Code tab inside the native Claude Desktop app. The app itself is a single container with multiple tabs (Chat, Cowork, Code). The Code tab is the one tuned for writing software.
The feature set mostly matches the terminal claude CLI, but GUI-only additions (visual diff, integrated terminal, drag-and-drop layout) change the feel. macOS shipped mid-2025. The Windows build went GA on 2026-02-10.
Naming, to save arguments later
- Claude Desktop — the full app, multiple tabs
- Claude Code Desktop — the Code tab inside that app
- Claude Code CLI — the standalone terminal binary
- Claude Code on the Web — the browser version at
claude.ai/code
Exec read. If your teams start arguing over "CLI vs. Web vs. Desktop," that's healthy. They're different shells over the same subscription. The real design question is who uses which surface, not which one wins.
2. Release history
Windows support (2026-02-10)
Anthropic shipped the Windows build of Claude Code Desktop on 2026-02-10. That moved the most advanced Anthropic agent onto roughly 70% of desktop users' machines overnight.
- Supported OS: Windows x64 and Windows ARM64 (separate installers)
- Prereq: Git for Windows
- Recommended: download directly from
claude.ai/download. The Microsoft Store build isn't recommended
The 2026-04-14 GA redesign
On 2026-04-14 Anthropic shipped a full redesign of the Code tab alongside a research preview of "Routines" (scheduled cloud automation). The redesign rethinks the GUI-first workflow.
| Feature | What it does |
|---|---|
| Multi-session sidebar | Run sessions against multiple repos in parallel, with sort and filter |
| Drag-and-drop layout | Chat, diff, terminal, editor, and preview panes can be arranged freely |
| Integrated terminal | Ctrl+` / Cmd+` opens a terminal inside the session (tests, scripts, etc.) |
| File editor | Small edits handled directly — replace, reformat, tidy |
| Visual diff | New diff viewer for large changes, with inline comments |
| Preview pane | HTML files, PDFs, dev-server output rendered inside the app |
| Side Chat (⌘+; / Ctrl+;) | Ask a side question without derailing the main thread |
| Streaming responses | Faster, steadier output |
Opus 4.7 support (2026-04-16)
- Released: 2026-04-16
- Client requirement: Claude Code v2.1.111 or newer
- Default Effort: Enterprise and Bedrock customers roll to Opus 4.7 automatically on 2026-04-23
- New command:
/ultrareview— simulates a senior reviewer and flags subtle design flaws
3. Surface comparison
Claude Code ships across five surfaces: Desktop, CLI, VS Code extension, JetBrains plugin, and Web. One subscription covers all of them. They're not interchangeable, though — each has a job it's best at.
| Capability | Desktop | CLI | VS Code | JetBrains | Web |
|---|---|---|---|---|---|
| Entry point | Native app (macOS / Win) | Terminal | Inside the IDE | Inside the IDE | Browser |
| Multi-session | Git worktree isolation | No | No | No | Cloud envs |
| Integrated terminal | Yes | Yes | No | No | Yes |
| Visual diff | Yes — diff pane | No | Yes — inline | No | Yes |
| Drag-and-drop layout | Yes | No | No | No | No |
| File editor | Yes | No | Yes | Yes | No |
| Local file execution | Yes | Yes | Yes | Yes | No |
| Remote sessions | No | Yes (SSH, /teleport) | No | No | Yes |
| Start from iOS | No | No | No | No | Yes — Dispatch |
| GitHub integration | Yes — PR watch | Yes | Yes | Yes | Yes |
| MCP connectors | Desktop Extensions | Yes | Yes | Yes | Yes |
| Scheduled runs (Routines) | No (local only) | Yes — /schedule | No | No | Yes |
| Subscription | Pro / Max / Team / Enterprise (shared across all surfaces) | ||||
Exec read. Desktop wins when you're sitting with the code and want conversation. Nightly batches and scheduled jobs live on Web or CLI via Routines. For non-engineer executives, Desktop is by far the gentlest start.
4. Feature inventory
Sessions and workflow
Parallel sessions. Each session gets its own Git worktree automatically. Nothing bleeds between them until you commit. Filter and group in the sidebar. Sessions can hand off to Web via /teleport, to an IDE, or to mobile.
Side Chat (⌘+; / Ctrl+;). Fire off a quick question without polluting the main thread. Good for spec lookups and "wait, how does X work" detours.
Permission modes
| Mode | Behavior |
|---|---|
| Ask permissions | Approves each file change (default) |
| Auto accept | Auto-approves for faster iteration |
| Plan mode | No file touches — just plans |
Code editing and review
- Visual diff — inline comments, per-line confirm
- File editor — direct edits for small fixes
- Review code — Claude reviews its own diff via the Review code button
- Adding context —
@filename, images, PDFs, drag-and-drop
Integrated tools
- Terminal — Ctrl+` to open; run tests, builds,
npmcommands - Preview — dev server, HTML, and PDF all preview inside the same pane
- PR watch — live status on GitHub PRs, CI results, auto-merge support
MCP and connectors
- Desktop Extensions (
.dxtfiles) — one-click install. Download, then drop into Settings → Extensions. No JSON editing. - MCP connectors — hundreds of servers available. MCP Tool Search uses lazy loading to cut context usage by up to 95%.
- Common ones: GitHub (PRs), Playwright (browser automation), Sentry (error tracking), Figma (design to code), Slack, Linear, Google Drive.
Plugins and skills
- Slash commands —
/model,/settings,/status,/schedule,/ultrareview, and more. Custom commands viaCLAUDE.mdwork across a team. - Hooks — pre/post execution hooks, configured in
settings.json/settings.local.json. - Subagents — multiple agents running different tasks in parallel, with a parent coordinating and merging.
Caveat. The official docs still leave some implementation details fuzzy — long-term storage of Side Chat, the signing and security model for Desktop Extensions, the learning logic behind Auto Memory. Run a PoC before committing.
5. Authentication and plan requirements
Free plan users don't get Claude Code at all. Pro is the floor. Routines (scheduled runs) scale by plan tier.
| Plan | Claude Code | Routines | Notes |
|---|---|---|---|
| Free | No | No | Chat only |
| Pro | Yes | 5 runs/day | $20/month |
| Max | Yes | 15 runs/day | $100–$200/month |
| Team | Yes | 25 runs/day | $150/seat/month, 5-seat minimum |
| Enterprise | Yes | Custom | Sales contract |
Team and Enterprise details
- Team Standard — Claude Code is not included. Needs a separate Team Premium upgrade.
- Team Premium — built for engineering teams. Claude Code and Routines included.
- Enterprise — SAML 2.0 / OIDC SSO, SCIM user provisioning, SOC 2 Type II, audit logs (30-day default retention), custom pricing.
6. File access and the permission model
Local sessions
- Full access inside whatever folder you opened in Desktop (usually a Git repo)
- Execution limited to the user's OS-level permissions
- Permission scheme — Ask / Auto accept / Plan mode
Remote sessions (Web / Cloud)
- GitHub auth — OAuth
- Repo restriction — Web sessions only run against GitHub-hosted repos (as of March 2026)
- Branch restrictions — by default, only branches prefixed
claude/can be pushed - Unrestricted push — opt in when creating a Routine
Enterprise / Bedrock route
- Zero Data Retention (ZDR) — Enterprise data isn't used for training
- BYOK — deploy via Amazon Bedrock, Google Vertex AI, or Microsoft Foundry
7. Enterprise considerations
Security and compliance
| Item | Spec |
|---|---|
| SSO | SAML 2.0, OIDC (Okta, Azure AD, etc.) |
| User provisioning | SCIM |
| Audit logs | Every user action, auth event, and file I/O tracked |
| Compliance | SOC 2 Type II |
| Encryption | TLS 1.3 in transit, AES-256 at rest |
| Data retention | 30 days default, customizable |
| Log export | JSON / CSV, push to Splunk, Datadog, Elastic |
Governance
- Central policy — tool permissions, file access restrictions at the org level
- Role-based access control
- Contribution dashboard — PRs generated, commit counts visible
What to know for Japan
Data residency. Claude Desktop and Cowork don't support Bedrock routing. That means the desktop app runs on Anthropic's US infrastructure. If your workload needs Japan-resident processing, the route is Claude Code CLI plus AWS Bedrock (Japan region).
Bedrock route.
- Opus 4.7 available in US East, Asia Pacific (Tokyo), Europe (Ireland), Europe (Stockholm)
- Geographic Inference Profiles lock processing to EU / US / APAC / Australia
- Setup:
CLAUDE_CODE_USE_BEDROCK=1,AWS_REGION=ap-northeast-1
8. Known limits and sharp edges
| Issue | Detail | Status |
|---|---|---|
| Rate limits burn fast | Max users' 5-hour sessions drained in 1–2 hours (March 2026) | Known — top priority |
| Prompt cache window | 5 minutes. Interrupt and you pay again | Design constraint |
| Opaque ceilings | Anthropic doesn't publish usage limits | Check the dashboard |
Service reliability
- 2026-04-15 incident — elevated errors on claude.ai, the API, and Claude Code. Login and usage tracking degraded.
Model support
- Opus 4.7 needs client v2.1.111 or newer
- Default Effort on Opus 4.7:
xhigh— watch the bill - Older models still work: Opus 4.6, Sonnet 4.6, Haiku 4.5
Windows specifics
- Git for Windows is required
- Don't use the Microsoft Store build — agent features aren't wired up
Exec read. The top complaint in the field is rate-limit unpredictability. For anything operationally critical, run multiple seats on Team or Enterprise plus a Bedrock fallback for burst traffic.
9. Ten exec scenarios for Desktop + Routines
All ten end in Slack notifications or draft PRs — the human decision point is obvious and the automation doesn't merge or publish.
- Daily code audit. Routine runs 09:00 — code review, security scan, style violations, complexity flags. Engineers stay focused on design work.
- Nightly regression watch. After CI passes, run smoke tests. Catches production issues early and pings Slack.
- Doc drift detection. Monday 10:00, it follows the week's merged PRs, checks API doc consistency, and opens update PRs.
- Cross-platform porting. A merge to the Python SDK triggers auto-ports to Go, Rust, and JavaScript SDKs.
- Support ticket to PR. Called from a CS tool, diagnoses the bug, opens a draft PR.
- Dependency audit. Thursday: npm / pip / Cargo scanned for CVEs, patch PRs proposed.
- Post-incident work. An issue tagged "incident" opens? Post-mortem drafted automatically.
- Regulatory sweep. Monthly scan for GDPR / PCI-DSS-relevant code, comments, and docs.
- Perf regression catch. After each benchmark run, latency / memory / CPU deltas go to Slack.
- Multi-region deploy check. Right after deploy, per-region health and localization checks.
Exec read. The pattern across all ten — the human decision stays at the end. Automate the diagnose / draft / notify steps. Merge, publish, and customer-facing sends are still human presses. That's the version of AI adoption that gets sign-off easily.
10. Pointers into the official docs
The full source list sits in the Sources block below. The three you'll actually open during rollout:
- Overview —
code.claude.com/docs/en/overview - Desktop quickstart —
code.claude.com/docs/en/desktop-quickstart - Enterprise / Bedrock —
claude.com/product/claude-code/enterpriseandcode.claude.com/docs/en/amazon-bedrock
This page reflects what was public on 2026-04-21. Anthropic ships fast; features and pricing change without notice.
Last verified: 2026-04-22
Sources
- Redesigning Claude Code on desktop for parallel agents — 2026-04-14 GA redesign (Anthropic Blog)
- Use Claude Code Desktop — Claude Code Docs
- Get started with the desktop app
- Introducing routines in Claude Code
- Claude Code overview
- Connect Claude Code to tools via MCP
- Claude Code for Enterprise
- Claude Code on Amazon Bedrock
- What's new in Claude Opus 4.7
- Claude pricing and plans