agent-manager

Product Introduction

1. Definition: Agent-manager is a terminal user interface (TUI) application written in Go that functions as a session manager and orchestrator for AI-powered coding agents (like Claude Code, OpenCode, and Codex) within a dedicated tmux server. It is a developer productivity tool that provides a unified dashboard for running, monitoring, and interacting with multiple command-line AI agents simultaneously.
2. Core Value Proposition: It exists to eliminate the terminal tab sprawl and context-switching overhead associated with managing multiple, concurrent AI coding sessions. Its primary value is enabling a fast, keyboard-driven workflow where developers can spawn, answer, and review the work of AI agents in seconds without leaving a single, centralized view, thereby accelerating iterative development and code review cycles.

Main Features

1. Unified Multi-Agent Dashboard: The TUI presents a live, foldable tree of projects and sessions, with each AI agent session displaying a real-time status (e.g., working, waiting, finished). This status is derived intelligently, for supported agents like Claude Code via MCP hooks, rather than simple text pattern matching. The view persists across manager restarts.
2. In-Context Agent Interaction: The space key serves a dual purpose: from the prompt bar, it instantly spawns a new agent with a task description; on an existing session row, it sends a reply directly into that agent's tmux pane. This allows users to unblock a waiting agent without manually attaching to its session, a key efficiency gain.
3. Integrated Whole-File Code Review: Pressing ctrl+r opens a dedicated, full-screen diff review mode for the selected agent's repository. It automatically detects the correct repo and target branch (often via MCP declarations from the agent itself), displays syntax-highlighted diffs, and allows users to add inline comments. These comments can be compiled into a single review prompt and sent back to the agent's session for iterative improvement.
4. Session Lifecycle & Orchestration: Sessions are independent tmux processes. Users can kill (x) a session to reclaim RAM while preserving its conversation history and identity, then later revive (v) it into the exact same thread. Conversations can be forked (f) into named sibling sessions. The manager also allows spawning plain shell terminals (T) alongside agents, all managed within the same hierarchy.
5. Extensible Tool Configuration: While it ships with built-in support for major AI CLI tools (Claude Code, OpenCode, etc.), it is extensible. Users can define custom tools in a config.toml file, specifying the launch command and regex-based rules for the manager to detect the tool's status (idle, working, errored), integrating any CLI into the managed workflow.

Problems Solved

1. Pain Point: Terminal and Context Management Overhead. Developers using multiple AI coding agents must manually manage numerous terminal tabs/windows, remember session states, and constantly cd into correct directories. This fragments focus and slows down the "task → agent → review" loop.
2. Target Audience: Full-stack and backend developers who iteratively use AI for coding tasks; tech leads reviewing AI-generated code across multiple features or bug fixes; open-source maintainers experimenting with AI assistants on different project branches. The primary persona is a keyboard-centric developer comfortable in the terminal using tools like tmux and git.
3. Use Cases: Parallel Feature Development: Spawning separate agents to work on different API endpoints simultaneously. Interactive Code Review: Using the diff review mode to guide an agent through refactoring a complex module. Experiment Forking: Using the conversation fork feature to explore alternative implementations without breaking the main agent's flow. Persistent Task Queues: Leaving long-running agents (e.g., writing tests) alive in a background tmux server while the laptop is closed.

Unique Advantages

1. Differentiation: Unlike standalone AI agent CLIs or IDE plugins, agent-manager does not replace the agents—it orchestrates them. It differs from other multi-session terminal managers by being purpose-built for the AI agent workflow, with deep features like status detection, in-context replying, and agent-aware code review. It is not a cost tracker or inter-agent communicator; it focuses purely on developer ergonomics and session control.
2. Key Innovation: The tight, bi-directional integration with the tmux layer and the agents via MCP. The manager runs its own MCP server that agents can call to self-declare state (like the active repository and target branch). This allows features like automatic repo detection for diffs and session auto-renaming, moving beyond simple terminal emulation to a cooperative management protocol.

Frequently Asked Questions (FAQ)

1. Does agent-manager replace Claude Code or other AI CLI tools? No, agent-manager is a session manager and orchestrator. It launches and manages the AI CLI tools you already have installed and configured, leveraging their full native functionality, logins, and MCP servers. It adds a management layer on top.
2. How does agent-manager handle git repositories and worktrees? It intelligently tracks the working directory for each session. Agents can declare their active repo via MCP. The alt+w shortcut spawns an agent directly into a new git worktree. The review mode (ctrl+r) automatically locates the correct repository and target branch for diff generation, prioritizing the agent's declarations.
3. What happens to my AI sessions if I quit the agent-manager application? All sessions continue running uninterrupted. Agent-manager uses a separate, dedicated tmux server (agentmgr). Quitting the TUI only closes the dashboard view; the underlying tmux sessions persist, allowing you to re-attach later or restart the manager to resume control.
4. Can I use agent-manager with my own custom or lesser-known AI CLI tool? Yes, through its extensible configuration. You can define a custom tool in the config.toml file, specifying its launch command and writing regex-based rules to detect its status (e.g., "working", "errored"), allowing it to integrate fully into the manager's status dashboard and workflow.
5. Is agent-manager available for Windows? Yes, but it requires the Windows Subsystem for Linux 2 (WSL2). It is a native Go binary that runs on macOS, Linux, and within a WSL2 Linux distribution, as it depends on Unix-like process and terminal management.