Yo. The noise in the AI coding agent space is getting deafening, but most tools still force you to juggle terminal tabs like a maniac. A new Show HN post introduces Agent Manager, a local terminal app designed to run multiple coding agents side by side without the context-switching nightmare. The project, hosted at github.com/wainmann/agent-manager, positions itself as a thin layer over existing CLIs, preserving your subscriptions, configs, and MCP servers while centralizing control.

The Architecture of Chaos

The core technical hook here is the use of persistent tmux sessions. Unlike ephemeral wrappers, Agent Manager launches your installed toolsβ€”Claude Code, Codex, OpenCode, Grok, Gemini CLI, Antigravity CLI, Pi, Command Code, Hermes Agent, Muse Code, Oh My Pi, Cursor CLI, and Kiro CLIβ€”as-is in isolated sessions. This means your login state, environment variables, and tool-specific features carry over exactly as they do in a plain terminal. The app groups these sessions into a project tree, allowing you to fold and reorder them. It’s not just a launcher; it’s a state manager. If a session dies, it revives on its own conversation. If you need to fork a conversation, you press 'f' and it spawns a separate named fork.

Keyboard-First and Mouse-Ready

While the pitch emphasizes a keyboard-first workflow, the implementation is surprisingly hybrid. The documentation details a robust set of keybindings: 'space' sends a prompt to the selected session, 'ctrl+r' opens a full-file diff with syntax highlighting, and 'c' allows you to comment on specific lines, which are then sent back to the agent as a review prompt. But don't let the keyboard-first label fool you; the mouse is fully integrated. You can drag rows to reorder them, click to focus, and use the mouse wheel to navigate. The interface even supports dragging the divider between the list and preview panes. It’s a pragmatic approach for those of us who live in the terminal but aren't afraid to click when it’s faster.

Ecosystem and Limitations

The tool currently supports a massive list of agents, including Hermes Agent, Muse Code, Cursor CLI, and Kiro CLI. Installation is straightforward via Homebrew on macOS/Linux or Scoop on Windows, with native support for psmux on Windows and WSL2. However, there is a glaring omission: cost tracking. The source explicitly states, "Not here yet: cost tracking." For developers running expensive LLMs across multiple concurrent agents, this is a significant gap. The project is Apache-2.0 licensed and actively seeking contributors, with a Discord community for setup help and workflow sharing.

Key Takeaways

  • Agent Manager uses persistent tmux sessions to isolate and manage multiple AI coding agents without losing state.
  • It supports 13+ major coding agents, including Claude Code, Codex, and Cursor CLI, preserving existing configs and MCP servers.
  • The interface offers deep keyboard controls for diff review and prompting, but also supports full mouse interaction for navigation.
  • Cost tracking is currently unavailable, a notable limitation for enterprise or high-volume users.
  • The project is open source (Apache-2.0) and actively seeking contributors and sponsors.

The Bottom Line

Agent Manager solves the orchestration problem for power users who are drowning in terminal tabs. While the lack of cost tracking is a pain, the ability to fork conversations and review diffs directly in the terminal is a killer feature for serious agentic workflows.