Claude Code capabilities

An artifact is a live, interactive web page that Claude Code publishes from a CLI (or Desktop app) session to a private URL on claude.ai, updating in place as the session continues. Introduced week of June 15-19, 2026 (v2.1.178-v2.1.183). Useful for PR walkthroughs with annotated diffs, dashboards built from session data, side-by-side option comparisons, and live investigation timelines, in place of pasting terminal output into chat tools.

Agent view lets you dispatch and monitor many Claude Code sessions in parallel from a single screen. Each background session is a full Claude Code conversation that keeps running without a terminal attached; you can come back, reply, and continue at any time. Research preview as of late 2025/early 2026.

Claude Code is built on top of the Claude Agent SDK. Anthropic exposes the same SDK (TypeScript and Python references) so developers can build custom agentic apps and workflows that share Claude Code's loop (tool use, sub-agents, MCP, skills, hooks). The GitHub Action is itself a thin wrapper around the SDK.

Lets you delegate coding tasks to Claude Code from inside Slack. When you mention @Claude in a channel with a coding request, Claude detects the intent, gathers context from the thread and recent channel messages, picks the relevant connected GitHub repo, and spins up a Claude Code session on the web. Claude posts progress updates back to the Slack thread and, when done, replies with a summary plus action buttons to View Session or Create PR. It runs under each user's own Claude account and counts against that user's plan rate limits.

Full Claude Code agent sessions that run in the browser on Anthropic-managed cloud infrastructure at claude.ai/code, with no terminal required. Each task runs in a fresh, isolated Anthropic-managed VM with your GitHub repo cloned; the session keeps running even if you close the browser and can be monitored from the Claude mobile app. Sessions can read GitHub issues, open and push to PRs, run tests, start services (PostgreSQL, Redis, Docker), and configure their own cloud environment via a setup script, network-access level, and environment variables. An Auto-fix mode lets Claude watch a PR and automatically respond to CI failures and review comments by pushing fixes to the branch. Sessions can be pulled into a local terminal with --teleport, and a new cloud session can be launched from the CLI with --remote. In research preview as of 2026.

Beyond the terminal/IDE, Claude Code is available as a web surface (research preview), with mobile access through the Claude mobile app rather than a separate Claude Code app. The web surface connects to your GitHub repos and runs sessions on Anthropic-managed cloud infrastructure; the mobile app lets you monitor, steer, and reply to cloud sessions (and enable Auto-fix on PRs). Web sessions can be resumed locally from the VS Code extension (Remote tab in Session history).

Introduced week of June 8-12, 2026 (v2.1.166-v2.1.176). When the primary model is overloaded, unavailable, or returns another non-retryable server error, Claude Code can automatically switch to a configured fallback model instead of failing the request, showing a notice when it switches. This is distinct from Fable 5's content-based safety-classifier fallback to Opus, which is a separate mechanism.

File editing is the agent's core loop: dedicated Read, Write, and Edit tools perform exact string replacements across single or multiple files; changes are previewed as inline diffs and require approval (or auto-accept) under the active permission mode. Plan mode lets you review and tweak the agent's plan before any code is touched. Checkpoints let you rewind code state to any prior message.

Official GitHub Action lets Claude Code respond to `@claude` mentions in issues and PRs, run on schedules, or be invoked as automation steps. It can implement features, create PRs, fix bugs, and answer questions, all following the repo's CLAUDE.md.

Hooks are user-defined handlers that run deterministically at specific points in Claude Code's lifecycle, letting you enforce policy, log activity, or block tool calls regardless of what the model decides. Session events: SessionStart, Setup, SessionEnd. Per-turn events: UserPromptSubmit, UserPromptExpansion, Stop, StopFailure. Agentic loop events: PreToolUse, PostToolUse, PostToolUseFailure, PostToolBatch, PermissionRequest, PermissionDenied. Subagent events: SubagentStart, SubagentStop. Compaction: PreCompact, PostCompact. Team/task events: TeammateIdle, TaskCreated, TaskCompleted. File/config events: InstructionsLoaded, ConfigChange, CwdChanged, FileChanged. Worktree events: WorktreeCreate, WorktreeRemove. MCP events: Elicitation, ElicitationResult. Display: Notification, MessageDisplay.

N/A. Claude Code is a terminal-first agentic CLI, not an inline code-completion engine. It does not provide AI ghost-text/tab completions as you type in an editor; instead, code changes flow through the agent's Read/Write/Edit tools and a diff-review surface. The CLI itself offers shell-style tab completion for command names and file paths, and the VS Code extension explicitly lists 'Tab completion' as a CLI-only feature that is NOT available in the graphical panel.

Claude Code's primary surface is an interactive terminal REPL launched with `claude`. You converse in natural language; the agent plans, reads files, edits them, runs shell commands, and reports back. Slash commands modify session state (clear, compact, model switch, plan mode, etc.) and `@`-mentions inject file contents.

Claude Code is available as a JetBrains plugin (beta) covering IntelliJ IDEA, PyCharm, Android Studio, WebStorm, PhpStorm, and GoLand. Brings IDE diff viewer, selection/active-tab context sharing, and Cmd/Ctrl+Esc launch.

Claude Code is an MCP host. Users add Model Context Protocol servers to connect external tools, databases, and APIs (e.g., GitHub, Slack, Google Drive, Asana). MCP servers are managed via CLI and the in-app `/mcp` dialog. The VS Code extension itself also runs a small built-in `ide` MCP server that the CLI connects to for native diff viewing, selection sharing, and Jupyter cell execution.

Claude Code lets users switch between Anthropic's frontier models per session: Sonnet (fast & capable, for the bulk of coding; the default on Pro and Team Standard seats), Opus (deep reasoning, for hard debugging / large refactors / architecture; the default on Max, Team Premium, Enterprise pay-as-you-go, and API accounts), Haiku (fastest/cheapest, for mechanical work), and Fable 5 (`/model fable`, a tier above Opus for the longest/hardest autonomous tasks — never the default on any plan). Mid-session switching preserves the conversation. A bundled `opusplan` mode plans with Opus and executes with Sonnet.

Claude Code is included in every paid consumer Claude plan (Pro, Max 5x, Max 20x), in Team and Enterprise, and is also accessible via direct API pay-as-you-go billing using an Anthropic Console key. Usage limits are shared between web Claude and Claude Code on the same subscription.

Plugins package slash commands, sub-agents, skills, MCP servers, and hooks into a single installable unit, distributed through plugin marketplaces (Git repos, URLs, or local paths). Installed plugins work across the terminal and the VS Code extension.

Claude Code automatically loads CLAUDE.md memory files as persistent project/user context on every turn, supports @-file mentions to inject file contents, and uses agentic Glob/Grep/Read tools to explore a codebase on demand. Memory layering (load order): managed/enterprise policy, then user, then project, then local — later files are read last.

Skills are organized folders containing a SKILL.md (YAML frontmatter + Markdown instructions, optional scripts and reference files). Claude loads them automatically when relevant, or the user invokes them with `/skill-name`. Custom commands have been merged into skills: files in `.claude/commands/*.md` still work, but `.claude/skills/<name>/SKILL.md` is the new recommended form. Follows the Agent Skills open standard (agentskills.io).

Sub-agents are specialized AI workers with their own isolated context window, system prompt, tool allowlist, model, and permissions. The orchestrator delegates side tasks (exploration, research, code review) to a sub-agent, which returns only a summary, preserving the main conversation's context.

Claude Code IS a terminal application. It can run arbitrary shell commands through its built-in Bash tool inside the user's working directory, subject to the active permission mode and per-command allow/deny rules. Sandboxing controls and `bypassPermissions` mode let advanced users run Claude with reduced friction in isolated environments.

Native VS Code extension provides a graphical Claude Code panel with inline diffs, plan-mode review, @-mentions with line ranges, session history, multiple parallel conversations in tabs/windows, checkpoints (rewind code), extended-thinking toggle, a GUI plugin/marketplace manager, and an optional terminal mode. The extension bundles its own copy of the CLI for the chat panel (it does not add `claude` to PATH — the standalone CLI install is still needed for terminal use). Same conversation history is shared with the CLI via `claude --resume`. Also installs on other VS Code forks (Cursor via a direct link, plus Devin Desktop, Kiro, etc.) through the Open VSX registry.

From v2.1.186 (week of June 22-26, 2026), `claude mcp login <name>` runs a configured remote MCP server's OAuth 2.0 flow directly from the shell, without opening the interactive `/mcp` panel inside a session. `claude mcp logout <name>` clears the stored credentials for that server. This closes a gap for non-interactive and scripted usage: previously OAuth for MCP servers could only be completed via the in-session `/mcp` panel.