github.devotd.
NewCoding agents

asheshgoplani/agent-deck v1.16.30 — Agent Session Management TUI

Introduction to asheshgoplani/agent-deck, a TUI repository that manages the terminal sessions of multiple AI coding agents with grouping, search, forking, worktree and cost tracking, and its latest release v1.16.30.

Repositoryasheshgoplani/agent-deck ↗

Introduction to asheshgoplani/agent-deck, a TUI repository that manages the terminal sessions of multiple AI coding agents with grouping, search, forking, worktree and cost tracking, and its latest release v1.16.30.

GitHub description: Terminal session manager for AI coding agents. One TUI for Claude, Gemini, OpenCode, Codex, and more.

stars
★ 1.1K
Forks
208
Language
Go
License
MIT
Latest release
v1.16.30
Read original source ↗View on GitHub ↗

asheshgoplani/agent-deck is a TUI (terminal user interface) tool that manages the terminal sessions of AI coding agents such as Claude Code, Gemini, OpenCode and Codex on a single screen. It is written in Go and released under the MIT license, and the latest release of the repository is v1.16.30. The changes contained in this release are not stated in the provided material.

Checking session status and switching between sessions with keyboard input in environments where multiple agents run at the same time is the center of this tool. The README introduces the tool as a mission control for handling agent sessions, and states that it supports session grouping and search, session forking, git worktree (working directory), and cost tracking. There is also a conductor feature that controls agent sessions from a mobile phone, and it can be used on macOS, Linux and Windows (WSL). The repository was created on December 3, 2025 and has 1128 stars.

The scores given by the research record are novelty 6.0, technology 7.0, quality 8.0 and momentum 3.9. Features and the internal module structure are broadly confirmed, but the assessment itself comes down to an incremental expansion within the session manager category. No performance measurements or benchmarks are presented, and although it can be called a choice that suits environments gathering multiple agent sessions on one screen, the range that can be expected does not go beyond the currently documented list of features.