A new interactive website called Claude Code Unpacked offers developers a visual tour of how Anthropic's command-line coding agent processes a request. The guide traces the path from a user's keyboard or piped input through the agent loop and ultimately to the rendered response, using a clickable source tree to organize its explanation.
The site's central walkthrough begins with the two supported input modes. Interactive text is captured through an Ink text-input component, while non-interactive operation reads data supplied through standard input. From there, visitors can follow the internal stages involved in handling a message instead of treating the tool as a single opaque prompt-and-response system.
Beyond that sequence, the guide catalogs more than 50 built-in tools available to the agent and groups them by purpose. It also organizes the program's slash commands, allowing visitors to inspect individual commands alongside relevant source references. This structure turns a large codebase into a navigable map for users interested in agent behavior, extension points or the mechanics behind familiar interface actions.
Claude Code Unpacked also highlights code paths that it describes as unreleased. The site labels these separately as feature-flagged, restricted by environment settings or commented out, an important distinction because their presence in source does not mean they are available in the public product. The examples shown include a terminal-based virtual pet whose characteristics derive from an account identifier and a persistent mode intended to consolidate memory between sessions and support autonomous background activity.
Other gated concepts in the guide include extended planning on Opus-class models with execution windows of as much as 30 minutes. It describes a lead-agent pattern that divides work among parallel workers in isolated Git worktrees, then gathers their outputs. Remote control from a phone or browser, including permission approvals, is another capability identified in the source map.
The site further points to background sessions launched with a command-line flag, using tmux beneath the interface, and communication between sessions over Unix-domain sockets. It describes a mechanism in which the system reviews completed work and organizes information learned across sessions. Again, these items are presented by the guide as code-discovered or gated features, not promises of release.
For developers, the project's value lies less in announcing a single feature than in making an agent application's moving parts easier to inspect. A visual map can help users understand where input enters, which tools an agent may invoke and how orchestration differs from a simple chat exchange. Because the supplied evidence comes from the guide itself, its interpretations should be treated as the author's reading of the source rather than official product documentation.
