Explore Claude Code

Compresses the current conversation into a shorter summary so Claude Code can keep working without hitting the context window limit.

Wipes the conversation history and starts a brand-new context, in the same working directory and session.

Analyzes the current project and generates a starting CLAUDE.md — the file Claude reads at the start of every session in that repo.

Opens the interface for creating, listing, and editing custom subagents — specialized configurations Claude can delegate focused work to.

Shows and lets you edit which tools and commands Claude is allowed to run automatically versus needing your approval for.

Switches which Claude model the current session uses — useful for trading off speed against depth of reasoning mid-task.

Installs the keybinding some terminals need for Shift+Enter to insert a newline instead of submitting the message.

Terminal launch flags (not in-session commands) for reopening a previous conversation instead of starting fresh.

Interrupts Claude mid-response or mid-tool-call — the fastest way to redirect without waiting for it to finish.

Pressing Esc twice opens a jump-back menu, letting you return to an earlier point in the conversation and continue differently from there.

Cycles Claude Code's permission mode, including toggling auto-accept edits mode on and off without going through /permissions.

Starting your input with ! runs the rest as a literal shell command, without Claude interpreting or reasoning about it first.

Starting your input with # saves it as a note to CLAUDE.md instead of sending it as a conversational message.

Cycles backward through your previously typed messages in the current session, the same way shell history works.

Clears whatever you've typed if the input line isn't empty; pressed again quickly on an empty line, it exits Claude Code.

Sends an end-of-file signal on an empty input line, exiting Claude Code — the same convention most shells use.

Inserts a line break in the input without submitting the message — plain Enter sends, Shift+Enter doesn't.

Completes a partially typed file path or slash command name, the same way shell tab-completion works.

Images can be added directly to a message by pasting from the clipboard or dragging a file into the input, without a separate upload step.

Clears the visible terminal screen, purely a display refresh — it doesn't touch the conversation history the way /clear does.

Toggles between a collapsed summary and the full detail of tool call output — useful when the short version hides something you need to see.

Shows every MCP server configured for the current session and whether each one is connected, still starting, or failed.

MCP servers can be registered per-project (shared via .mcp.json) or per-user (available in every project you open).

The CLI command for registering a new MCP server, specifying its start command and which scope it should apply to.

A server that needs to download dependencies or initialize a heavier process on first launch can take noticeably longer than one already warmed up.

MCP servers communicate over a specific transport — commonly stdio for local processes, or HTTP/SSE for remote ones — and the config has to match what the server actually speaks.

Many MCP servers need their own API key or config value passed in as an environment variable — without it, they fail to start or authenticate.

Tools exposed by an MCP server go through the same permission system as any other tool — the first use of a new one typically prompts for approval.

Some MCP servers authenticate via a browser-based OAuth flow rather than a static API key, similar to how Claude Code's own login works.

A JSON syntax error anywhere in .mcp.json can prevent the whole file from parsing, taking every server in it down at once, not just a specific entry.

A server configured to run via npx has to successfully resolve and download that package before it can even start — network or registry issues surface here first.

The conversation, including file contents and tool output, has grown too large for the model to hold in a single request.

An informational notice, not an error — it just means this project doesn't have a CLAUDE.md file with project-specific guidance yet.

A tool call was stopped by a permission rule or hook, rather than by Claude itself deciding not to run it.

The Anthropic API returned a temporary rate-limit or capacity error — usually resolved automatically with a short retry.

Claude Code can't find a custom command you expected to exist — almost always a file location or naming issue.

A configured MCP (Model Context Protocol) server couldn't start or connect, so its tools aren't available this session.

Claude Code needs an interactive terminal to work normally — this shows up when it's run somewhere that isn't one, like piped input or certain CI environments.

Claude Code's credentials — whether from a login flow or an API key — have expired, been revoked, or aren't configured, and it can't authenticate.

A syntax error in settings.json can prevent the whole file from being read, silently dropping permissions, hooks, and other configuration defined in it.

A hook defined in settings.json isn't firing when expected — almost always a mismatch between the event name, matcher, or an executable permission issue.

Very large files get truncated when read, since the full content would consume an outsized share of the context window on its own.

Git commands go through the same permission system as any shell command — without a specific rule, each one (or each new git subcommand) can prompt individually.

Path-related errors on Windows are usually a mismatch between Windows-style and Unix-style paths, especially when WSL is involved.

A corporate network proxy that intercepts outbound HTTPS can block Claude Code's API connection unless it's explicitly configured to route through it.

A session using far more tokens than expected is usually explained by large file reads, verbose tool output, or a long conversation that hasn't been compacted.