ACP Client 'agent-shell' for Emacs Released

The package agent-shell has been released, allowing integrated control of various coding agents such as Claude Code, Codex, and Gemini CLI via ACP (Agent Client Protocol) in Emacs. Unlike traditional terminal emulator approaches (such as vterm), it utilizes Emacs' built-in comint-mode buffers, ensuring that the same UI and operation methods are maintained even when switching agents.

Key Features and Architecture

• Integrated Control: Uses Emacs' standard editing, search, and navigation commands directly, and allows sharing of MCP server configurations.
• Viewport: Opened with C-c C-o, it allows collapsing tool calls and reasoning processes into Activity groups (available from version 0.63).
• Steering Support: Through ACP extensions, Steering functionality, which inserts prompts into ongoing turns, is supported by the Claude and Codex adapters.
• Permission Management: Changes can be reviewed in a diff buffer, and users can navigate to the file location using RET.

Supported Agents and Installation

The official support list includes a total of 19 agents, such as Claude Code, Codex, Gemini CLI, Goose, and Cursor. While Gemini CLI and Goose use built-in ACP modes, Claude Code and Codex require the installation of separate npm adapters (such as @agentclientprotocol/claude-agent-acp). It is distributed via MELPA and follows the GPL-3.0 license.