Claude surfaces
The places you can use Claude: Claude.ai, Claude Desktop, Claude Code (terminal, IDE, and web), Claude Cowork, and the Claude API.
What it is
Ways to talk to the same Anthropic models. Each surface has different strengths and different component support. Pick the surface that matches your work, then read its sub-page for install and feature detail.
| Surface | Best for | Components supported |
|---|---|---|
| Claude.ai | Chat, document upload, one-click Connectors | Connectors, Projects |
| Claude Desktop | The same chat plus local files via MCP | Connectors, MCP servers |
| Claude Code | Multi-step coding (CLI, IDE, web) | Skills, MCP servers, Plugins, Hooks, Routines |
| Claude Cowork | Non-coding desktop automation in a sandboxed VM | Connectors, Plugins |
| Claude API | Building your own product, agent, or backend | SDKs, Managed Agents, MCP |
Your account, Projects, and conversation history are shared across Claude.ai and Claude Desktop. Claude Code and Cowork are billed through your Pro/Max/Team/Enterprise subscription or via API.
Cross-cutting features
Some features span more than one surface — you start them in one place and view or extend them in another. Read the linked page for detail.
- Routines — scheduled remote Claude Code agents. Created via the
/scheduleskill in Claude Code, viewed and edited atclaude.ai/code/routines(a Claude.ai surface), execute on Anthropic cloud infrastructure with MCP connectors attached. - Background sessions —
claude --bgstarts a long-running Claude Code task; reattach later via/resumeorclaude agents. Spans CLI andclaude.ai/code. - MCP tunnels (Research Preview) — outbound-only
cloudflaredtunnels that expose private MCP servers to Claude.ai. Bridges Claude.ai / Claude Desktop and MCP servers running inside private networks. See MCP servers. - Channels (Research Preview) — inverts MCP from pull to push: an allowlisted MCP server can inject events into a running Claude Code session so Claude reacts to outside happenings (CI failures, monitoring alerts, an iMessage). Start a session with
claude --channels discord,telegram,imessage; bundled plugins are Telegram, Discord, iMessage, andfakechat. Requires aclaude.ailogin (no API-key auth); Team / Enterprise admins must setchannelsEnabled: trueand curateallowedChannelPlugins. Custom channels need--dangerously-load-development-channels. Bridges Claude Code and MCP servers. Shipped 2026-03-20 in v2.1.80. - Claude Security (Enterprise public beta) — vulnerability scanning at
claude.ai/security(sidebar icon in Claude.ai). Launched 2026-05-04 on Opus 4.7. Each finding has a remediation button that opens a Claude Code on the web session, drafts a patch, and opens a PR. Enterprise plans only; Team / Max coming. - Dynamic Workflows (Research Preview) — spans CLI, Desktop, and the VS Code extension of Claude Code. Claude writes an orchestration script on the fly and runs up to 1,000 parallel subagents (16 concurrent) to tackle migrations, audits, and other large jobs. Trigger one task by including the keyword
ultracodein a prompt (renamed fromworkflowin v2.1.160 — the bare wordworkflowno longer fires), or set/effort ultracodeto auto-orchestrate every substantive task in the session. Launched 2026-05-28 alongside Opus 4.8; requires Claude Code v2.1.154+. On by default on Max and Team; admin-gated on Enterprise; off by default on Pro (toggle in/config). Burns substantially more tokens than a normal session. - Sandboxing differs by surface — useful to know when reasoning about what Claude can touch on your machine. Per How we contain Claude across products (Anthropic engineering, late May 2026): Claude.ai runs tools inside gVisor; Claude Code running locally uses macOS Seatbelt or Linux Bubblewrap (reads allowed, writes inside the workspace, network denied by default); Claude Cowork runs a full VM (Apple Virtualization on macOS, HCS on Windows). Local MCP servers you wire into Claude Desktop run with your user’s permissions, not inside a sandbox — review folders and scripts before exposing them.
Common pitfalls
- Confusing Claude Desktop, Claude Code, and Claude Cowork — three different apps that all live near
claude.ai. - Components don’t all install everywhere. Skills, MCP servers, and Plugins work in Claude Code and Cowork; Connectors work in Claude.ai / Desktop / Cowork. A few tools (e.g., PubMed) ship in multiple places.
- API usage is metered separately from a Pro subscription.
See also
- Decision tree
- Skills, MCP servers, Plugins, Connectors
- Claude Code overview — canonical docs
- Claude.ai help center — canonical docs
Sources
- Claude Code product landing — Anthropic product page; verified 2026-05-26 (this run) — canonical install command and Windows/WinGet/Linux package options.
- Set up Claude Code — Anthropic docs; verified 2026-05-26 (this run).
- Claude Code on the web — Anthropic docs; verified 2026-05-19.
- Claude Code on the web (announcement) — published 2025-10-20.
- Anthropic announces Claude Cowork — published 2026-01-13.
- Cowork and plugins for teams across the enterprise — Anthropic blog; published 2026-02-24.
- Claude Code changelog (v2.1.144 / v2.1.145, May 18–19 2026) —
claude --bg,/resume,claude agents --json; verified 2026-05-26 (this run). - Claude Security is now in public beta — Anthropic blog; published 2026-05-04 —
claude.ai/security, Opus 4.7, remediation opens Claude Code on the web, Enterprise public beta. - Getting started with Claude Security — Anthropic tutorial; verified 2026-05-26 — sidebar location and remediation-session flow.
- Introducing dynamic workflows in Claude Code — Anthropic blog; published 2026-05-28 — research preview, 1,000-agent cap, 16 concurrent, plan availability and defaults.
- Orchestrate subagents at scale with dynamic workflows — Anthropic docs; verified 2026-06-06 (this run) — single-task keyword renamed
workflow→ultracodein v2.1.160 (bareworkflowno longer triggers);/effort ultracodeis a session setting (sends xhigh + auto-orchestration), not a model effort level; effort ladder remains low/medium/high/xhigh/max; v2.1.154 requirement. - Introducing Claude Opus 4.8 — Anthropic news; published 2026-05-28 — Opus 4.8 ship date, paired Dynamic Workflows + effort-control launch.
- How we contain Claude across products — Anthropic engineering blog; published late May 2026 (verified 2026-05-31 this run) — per-surface sandbox stack: gVisor for Claude.ai, Seatbelt (macOS) / Bubblewrap (Linux) for Claude Code local with default-deny network, full VM (Apple Virtualization / HCS) for Cowork.
- How we contain Claude across products (link blog) — Simon Willison; published 2026-05-30 — independent summary of the same engineering post.
- Channels reference — Anthropic docs; verified 2026-06-03 (this run) —
--channels, bundled plugins,claude.aiOAuth-only auth, Team / EnterprisechannelsEnabled+allowedChannelPluginsmanaged settings,--dangerously-load-development-channels. - Claude Code 2.1.80: –channels Lets MCP Servers Push Messages Into Your Session — Vibe Sparking AI; published 2026-03-20 — research-preview launch context, Telegram / Discord initial channels.