Detailed Analysis
A Reddit user has released a community-built terminal user interface (TUI) designed to restore and improve upon functionality that Anthropic removed from Claude Code: the "/agents" wizard, a feature that previously allowed developers to manage custom subagents directly within the CLI tool. The replacement tool, built independently and distributed via GitHub, extends the original concept significantly by enabling cross-project agent management through bookmarks, global agent configuration (stored in .claude/users), and visibility into plugin-provided agents housed in .claude/plugins/marketplaces. Notably, the tool distinguishes between editable user-created agents and read-only plugin agents, a safeguard against accidentally breaking third-party plugin functionality, while still offering an explicit "tracking" mechanism for developers who want to intentionally modify plugin agent behavior for development or customization purposes.
The development is notable less for its novelty as software and more for what it reveals about the Claude Code ecosystem's maturation. Claude Code, Anthropic's agentic coding CLI, has increasingly become a platform on which third-party tooling is built rather than merely a closed product. The removal of the native /agents wizard evidently left a gap that at least one power user felt compelled to fill, and the resulting tool—cross-platform across macOS, Linux, and Windows, with update mechanisms via git pull and integration with default text editors like vim, nano, or VS Code—reflects the kind of infrastructure-layer tooling typically seen around mature developer platforms like Docker or Kubernetes, not just a single AI assistant.
Equally significant is the meta-narrative embedded in the post: the tool itself was built almost entirely through "vibe coding" using Claude Desktop's Cowork feature paired with Sonnet, with the human author doing only light auditing for syntax and comment cleanup. The creator's workflow—breaking the project into discrete tasks and issuing one prompt per task rather than requesting a monolithic output—illustrates an emerging best practice among practitioners for managing agentic coding sessions, where decomposition improves reliability over single massive requests. The offhand admission of accidentally running with extended thinking mode enabled, consuming 90% of a weekly usage allocation, also surfaces a practical friction point that Claude Pro and Max subscribers regularly navigate: balancing model reasoning depth against consumption limits within Anthropic's tiered usage caps.
Broadly, this episode fits into a wider pattern in which end users are increasingly building and open-sourcing plugins, wizards, and management layers atop Anthropic's developer tools rather than waiting for official features. This grassroots tooling ecosystem mirrors dynamics seen in other extensible platforms, where the vendor provides core primitives (subagents, plugins, marketplaces) and the community fills usability gaps. It also underscores how capable current-generation models like Sonnet have become at autonomously scaffolding functional developer tools with minimal human intervention beyond task decomposition and light review—an early but tangible sign of AI systems increasingly building the tooling that supports other AI systems.
Read original article →