Agent Shell

Agent Shell is an Emacs package that gives you a native shell buffer for talking to coding agents over the Agent Client Protocol (ACP). It builds on the built-in comint shell through shell-maker and uses acp.el for the protocol, so any ACP-speaking agent can be driven from inside the editor.

Configurations ship for Claude Agent, Codex, Gemini CLI, Goose, Grok Build, Cursor, Kimi Code CLI, CodeBuddy, Kiro CLI, Qwen Code, Auggie, Mistral Vibe, Factory Droid, Pi, Oh My Pi, Opencode, and Antigravity. M-x agent-shell starts or reuses a session for any configured agent, and a prefix argument forces a new one so several shells can run at once. Authentication is per provider and accepts a login flow, an API key, an OAuth token, Vertex AI or enterprise variants, or credentials managed outside Emacs.

Alongside the shell buffer, a viewport presents one request and response at a time, splitting composing and reading into separate major modes with their own key maps in a buffer named after the shell. A prompt sent while the agent is working is queued and submitted when the turn ends. Agents advertising the _session/steering ACP extension can instead take it into the running turn, which the Claude and Codex packages implement; steering is always explicit, replaces rather than adds to what the agent was doing, and a declined steer interrupts the turn and says so rather than letting the agent continue in a direction you believe you corrected.

Agents can be run inside Docker containers or devcontainers by prefixing the launch command, either with a static list or a function that picks a container per buffer or project. A path resolver maps container paths to the local working directory, and turning off the text-file client capabilities keeps the agent reading and writing inside the container instead of on the host filesystem. MCP servers are declared once in agent-shell-mcp-servers rather than repeated per agent.

Session controls surface in the shell header once the agent advertises them: model, session mode, and thought level, each clickable and bound to a key. Per-project data such as transcripts and screenshots goes to a .agent-shell/ directory in the project root, added to .gitignore the first time it is created. Every setting can be applied per project through directory-local variables.