#+TITLE: Emacs Agent Shell #+AUTHOR: Álvaro Ramírez

[[https://melpa.org/#/agent-shell][file:https://melpa.org/packages/agent-shell-badge.svg]]

👉 [[https://github.com/sponsors/xenodium][Support this work via GitHub Sponsors]] by [[https://github.com/xenodium][@xenodium]] (check out my [[https://xenodium.com][blog]])

[[file:agent-shell.png]]

This project needs your funding

As you pay for those useful LLM tokens, consider [[https://github.com/sponsors/xenodium][sponsoring]] development and maintenance of this project. With your help, I can make this effort more [[https://github.com/sponsors/xenodium][sustainable]].

Thank you!

[[https://xenodium.com/][Alvaro]]

agent-shell

A native Emacs shell to interact with LLM agents powered by ACP ([[https://agentclientprotocol.com][Agent Client Protocol]]).

With agent-shell, you can interact with any ACP-driven agent like:

[[https://code.claude.com/docs/en/overview][Claude Agent]] (Anthropic)
[[https://github.com/agentclientprotocol/codex-acp][Codex]] (OpenAI)
[[https://github.com/google-gemini/gemini-cli][Gemini CLI]] (Google)
[[https://goose-docs.ai/docs/getting-started/installation][Goose]]
[[https://docs.x.ai/docs/overview][Grok Build]] (xAI)
[[https://cursor.com/docs/cli][Cursor]]
[[https://www.kimi.com/code][Kimi Code CLI]]
[[https://www.codebuddy.ai][CodeBuddy]]
[[https://kiro.dev][Kiro CLI]]
[[https://github.com/QwenLM/qwen-code][Qwen Code]]
[[https://docs.augmentcode.com/cli/overview][Auggie]]
[[https://github.com/mistralai/mistral-vibe][Mistral Vibe]]
[[https://factory.ai][Factory Droid]]
[[https://github.com/svkozak/pi-acp][Pi]]
[[https://github.com/can1357/oh-my-pi][Oh My Pi]]
[[https://opencode.ai][Opencode]]

[[https://www.youtube.com/watch?v=R2Ucr3amgGg][agent-shell - YouTube]]

[[https://www.youtube.com/watch?v=R2Ucr3amgGg][file:yt.png]]

[[https://www.youtube.com/watch?v=ymMlftdGx4I][agent-shell + Claude Skills - YouTube]]

[[https://www.youtube.com/watch?v=ymMlftdGx4I][https://img.youtube.com/vi/ymMlftdGx4I/0.jpg]]

[[https://www.youtube.com/watch?v=HJQ86HuSIJI][agent-shell + Claude Skills + Charts - YouTube]]

[[https://www.youtube.com/watch?v=HJQ86HuSIJI][https://img.youtube.com/vi/HJQ86HuSIJI/0.jpg]]

[[https://www.youtube.com/watch?v=7fdHeUoRRgg][Bending Emacs Episode 14: Prototyping iOS apps with agent-shell, artist-mode, and Claude Skills - YouTube]]

[[https://www.youtube.com/watch?v=7fdHeUoRRgg][https://img.youtube.com/vi/7fdHeUoRRgg/0.jpg]]

News

Do take a look at all these blog posts as they showcase most =agent-shell= features.

[[https://xenodium.com/agent-shell-0-55-updates][agent-shell 0.55 updates]].
[[https://xenodium.com/agent-shell-0-47-1-updates][agent-shell 0.47 updates]].
[[https://xenodium.com/agent-shell-0-25-updates][agent-shell 0.25 updates]].
[[https://xenodium.com/agent-shell-016-improvements-melpa][agent-shell 0.17 improvements + MELPA]].
[[https://xenodium.com/agent-shell-0-5-improvements][agent-shell 0.5 improvements]].
[[https://xenodium.com/introducing-agent-shell][Introducing Emacs agent-shell (powered by ACP)]].
[[https://xenodium.com/introducing-acpel][Introducing acp.el]].

Related projects

=agent-shell= relies on [[https://github.com/xenodium/acp.el][acp.el]] to communicate with agents via ACP ([[https://agentclientprotocol.com/][Agent Client Protocol]]).

We now have a handful of additional packages to extend the =agent-shell= experience:

[[https://github.com/xenodium/emacs-skills][emacs-skills]]: Claude Agent skills for Emacs.
[[https://github.com/ElleNajt/agent-shell-to-go][agent-shell-to-go]]: Interact with =agent-shell= sessions from your mobile or any other device via Slack.
[[https://github.com/Embedded-Focus/agent-circus][agent-circus]]: Run AI coding agents in sandboxed Docker containers.
[[https://github.com/cmacrae/agent-shell-sidebar][agent-shell-sidebar]]: A sidebar add-on for =agent-shell=.
[[https://github.com/dcluna/agent-shell-bookmark][agent-shell-bookmark]]: Bookmark support for agent-shell sessions.
[[https://github.com/gveres/agent-shell-workspace][agent-shell-workspace]]: Dedicated tab-bar workspace for managing multiple =agent-shell= sessions.
[[https://github.com/jethrokuan/agent-shell-manager][agent-shell-manager]]: Tabulated view and management of =agent-shell= buffers.
[[https://github.com/nineluj/agent-review][agent-review]]: Code review interface for =agent-shell=.
[[https://github.com/ultronozm/agent-shell-attention.el][agent-shell-attention.el]]: Mode-line attention tracker for =agent-shell=.
[[https://github.com/xenodium/agent-shell-knockknock][agent-shell-knockknock]]: Notifications for =agent-shell= via [[https://github.com/konrad1977/knockknock][knockknock.el]].
[[https://github.com/zackattackz/agent-shell-notifications][agent-shell-notifications]]: Desktop notifications for =agent-shell= events.
[[https://github.com/ElleNajt/meta-agent-shell][meta-agent-shell]]: Multi-agent coordination system for =agent-shell= with inter-agent communication, task tracking, and project-level dispatching.
[[https://github.com/cxa/agent-shell-macext][agent-shell-macext]]: macOS-specific enhancements for =agent-shell=.
[[https://github.com/lllShamanlll/agent-shell-org-transcript][agent-shell-org-transcript]]: Org-mode transcripts for =agent-shell= sessions, with org-roam integration.
[[https://github.com/eddof13/ob-agent-shell][ob-agent-shell]]: Org Babel backend for executing =agent-shell= source blocks.
[[https://github.com/Marx-A00/agent-recall][agent-recall]]: Search, browse, and resume =agent-shell= conversation transcripts.
[[https://github.com/lgmoneda/agent-shell-pet][agent-shell-pet]]: Codex-like pets that broadcast agent-shell session states.
[[https://github.com/alberti42/agent-shell-math-renderer][agent-shell-math-renderer]]: Render LaTeX math equations as SVGs in =agent-shell=.
[[https://github.com/junyi-hou/agent-shell-tramp][agent-shell-tramp]]: Tramp integration for =agent-shell=.
[[https://github.com/nohzafk/agent-shell-hud][agent-shell-hud]]: Real-time =agent-shell= status overlay via a floating dashboard.
[[https://github.com/timfel/agent-shell-desktop.el][agent-shell-desktop]]: Desktop save mode integration for =agent-shell=. Saves and resumes shells by folder, config, and session id.
[[https://github.com/ultronozm/agent-shell-links.el][agent-shell-links.el]]: Bookmarks and Org links for =agent-shell= sessions.
[[https://github.com/wandersoncferreira/agent-shell-dashboard][agent-shell-dashboard]]: A landing page for =agent-shell=.

Articles

20Y: [[https://20y.hu/~slink/journal/agent-shell/index.html][Agentic development workflow in Emacs]]
Skye: [[https://skyefreeman.com/blog/2026/04/15/how-im-using-llms-with-emacs][How I'm using LLM's via Emacs]].
sincebyte: [[https://neoemacs.com/posts/agent-shell-cursor-agent-cli/][Agent Shell Cursor Agent Cli]].
Naputo: [[https://blog.n-daisuke897.com/posts/2026-03-01-agent-shell-session-selection/][How to Select Sessions When Starting the agent-shell ACP Client Running on Emacs]]
vandee: [[https://www.vandee.art/blog/2026-01-23-my-agent-practice-with-opencode-in-emacs.html][My Agent Practice with OpenCode in Emacs]].
Fedora Project Wiki: [[https://fedoraproject.org/wiki/SIGs/AI-ML#agent-shell][AI/ML Special Interest Group]].

Icons

Thanks to [[https://github.com/lobehub/lobe-icons][Lobe Icons]] for the lovely icons.

Setup

** External dependencies

*** Claude Agent SDK

For Anthropic's [[https://code.claude.com/docs/en/overview][Claude Agent]] (formerly known as the Claude Code), follow [[https://github.com/agentclientprotocol/claude-agent-acp][claude-agent-acp instructions]], typically something like:

#+begin_src bash npm install -g @agentclientprotocol/claude-agent-acp #+end_src

Note: The =-g= flag is required to install the binary globally so it's available in your PATH. After installation, verify it's available by running =which claude-agent-acp= in your terminal.

Optionally: You may also need [[https://code.claude.com/docs/en/overview][Claude Agent]] itself if you want to use a Claude subscription (run the CLI outside of Emacs at least once to log in to your subscription and then use =agent-shell= from Emacs). Follow Claude Agent's [[https://code.claude.com/docs/en/overview#get-started][get started]] for installation.

*** Codex

For OpenAI's Codex, install [[https://github.com/agentclientprotocol/codex-acp][@agentclientprotocol/codex-acp]] (=npm install -g @agentclientprotocol/codex-acp=) and ensure the =codex-acp= executable is in PATH.

*** Gemini CLI

For Google's [[https://github.com/google-gemini/gemini-cli][Gemini CLI]], be sure to get a recent release supporting the =--experimental-acp= flag.

*** Goose

For Goose CLI, install [[https://goose-docs.ai/docs/getting-started/installation][goose]] and ensure the =goose= executable is in PATH.

*** Grok Build (xAI)

For [[https://docs.x.ai/docs/overview][Grok Build]], install the Grok CLI so =grok= is on =PATH= (the installer typically puts it at =~~/.grok/bin=). Authenticate once via the CLI login flow (credentials live in =~~/.grok/auth.json=).

Grok Build speaks ACP over stdio:

#+begin_src bash grok agent stdio #+end_src

Optional flags (set via =agent-shell-xai-acp-command=):

#+begin_src bash grok agent -m grok-build stdio grok agent --always-approve stdio #+end_src

See also the [[https://zed.dev/acp/agent/grok-build][Zed ACP registry listing]] for Grok Build.

*** Cursor

For Cursor agent, install the official [[https://cursor.com/docs/cli][Cursor CLI]] and ensure the =agent= executable is in PATH. =agent-shell= uses its official ACP server (=agent acp=).

A recent Cursor CLI is required (the =acp= subcommand was added in 2026). Older CLIs won't expose =agent acp=.

*** Kimi Code CLI

For Kimi Code CLI, install with:

#+begin_src bash curl -L code.kimi.com/install.sh | bash #+end_src

See https://www.kimi.com/code for details.

*** CodeBuddy

For CodeBuddy agent, install the CodeBuddy CLI and ensure ACP mode is available. For example:

#+begin_src bash codebuddy --acp #+end_src

See https://www.codebuddy.ai/docs/zh/ide/Getting-Started/Installation for details.

*** Kiro CLI

For Kiro CLI, install with:

#+begin_src bash curl -fsSL https://cli.kiro.dev/install | bash #+end_src

See https://kiro.dev/docs/cli/acp/ for details.

*** Qwen Code

For Qwen Code, install with:

#+begin_src bash npm install -g @qwen-code/qwen-code@latest #+end_src

See https://github.com/QwenLM/qwen-code for details.

*** Auggie

For Auggie CLI, install with:

#+begin_src bash npm install -g @augmentcode/auggie #+end_src

See https://docs.augmentcode.com/cli/overview for details. *** Mistral Vibe

For Mistral Vibe, install with:

#+begin_src bash uv tool install mistral-vibe #+end_src

See https://github.com/mistralai/mistral-vibe for details.

*** Factory Droid

For Factory Droid, install Droid:

#+begin_src bash curl -fsSL https://app.factory.ai/cli | sh #+end_src

See https://factory.ai for details.

*** Pi

For Pi coding agent, install the =pi-acp= adapter:

#+begin_src bash npm install -g pi-acp #+end_src

See https://github.com/svkozak/pi-acp for details. *** Oh My Pi (=omp=)

For the =omp= agent, install with:

#+begin_src bash bun install -g @oh-my-pi/pi-coding-agent #+end_src

See https://github.com/can1357/oh-my-pi for details.

*** Opencode + local ollama models

For opencode coding agent, install with:

#+begin_src bash curl -fsSL https://opencode.ai/install | bash #+end_src

See https://opencode.ai/ for details.

For local ollama models, sudo permission is needed to initialize ~~ollama.service~~. The following will install ollama and a local Qwen2.5:7b model.

#+begin_src bash curl -fsSL https://ollama.com/install.sh | sh ollama pull qwen2.5:7b #+end_src

If you are behind a network proxy, ~~ollama pull~~ will timeout. You then need to add proxy settings inside ~~/etc/systemd/system/ollama.service~~

See https://ollama.com/ for details

** Installation

=agent-shell= is powered by built-in =comint-shell=, via [[https://github.com/xenodium/shell-maker][shell-maker]], available on [[https://melpa.org/#/shell-maker][MELPA]].

Both [[https://melpa.org/#/agent-shell][agent-shell]] and its dependency [[https://melpa.org/#/acp][acp.el]] are now available on MELPA.

You can install via:

#+begin_src emacs-lisp (use-package agent-shell :ensure t :ensure-system-package ;; Add agent installation configs here ((claude . "brew install claude-code") (claude-agent-acp . "npm install -g @agentclientprotocol/claude-agent-acp"))) #+end_src

This will automatically install the required dependencies ([[https://melpa.org/#/acp][acp.el]] and [[https://melpa.org/#/shell-maker][shell-maker]]).

*** Optional: image utilities

Pasting images from the clipboard (=agent-shell-send-clipboard-image= or =yank= or =C-y=) and capturing screenshots (=agent-shell-send-screenshot=) rely on external tools. Install the ones for your platform:

| Platform | Clipboard paste | Screenshot capture | |-----------------+------------------------------------------+------------------------------| | macOS | =brew install pngpaste= | =screencapture= (preinstalled) | | Linux (Wayland) | =wl-paste= (e.g. =apt install wl-clipboard=) | =import= (from ImageMagick) | | Linux (X11) | =apt install xclip= | =import= (from ImageMagick) | | Windows | =powershell= | (not bundled) |

Handlers are detected at runtime via =executable-find=. See =agent-shell-clipboard-image-handlers= and =agent-shell-screenshot-command= to customize.

*** Doom Emacs

If you are using Doom Emacs and would like to use the =package!= macro:

#+begin_src emacs-lisp (package! shell-maker) (package! acp) (package! agent-shell) #+end_src

Run =doom sync= and restart.

Include =require= before configuration:

#+begin_src emacs-lisp (require 'acp) (require 'agent-shell) ;; rest of config... #+end_src

** Configuration

Configure authentication for the agent providers you want to use.

*** Environment variables

Pass environment variables to the spawned agent process by customizing the agent-shell-*-environment variable with agent-shell-make-environment-variables. The helper accepts key/value pairs and exports them when the agent starts.

#+begin_src emacs-lisp (setq agent-shell-anthropic-claude-environment (agent-shell-make-environment-variables "ANTHROPIC_API_KEY" (auth-source-pass-get 'secret "anthropic-api-key") "HTTPS_PROXY" "http://proxy.example.com:8080")) #+end_src

**** Inheriting environment variables

By default, the agent process starts with a minimal environment. To inherit environment variables from the parent Emacs process, use the :inherit-env t parameter in agent-shell-make-environment-variables:

#+begin_src emacs-lisp (setenv "ANTHROPIC_API_KEY" (auth-source-pass-get 'secret "anthropic-api-key"))

(setq agent-shell-anthropic-claude-environment (agent-shell-make-environment-variables :inherit-env t)) #+end_src

This ensures that environment variables like PATH, HOME, and others from your Emacs session are available to the agent process, while still allowing you to override or add specific variables.

**** Loading environment variables from files

You can load environment variables from .env files using the :load-env parameter. This supports both single and multiple files:

#+begin_src emacs-lisp ;; Load from a single .env file (setq agent-shell-anthropic-claude-environment (agent-shell-make-environment-variables :load-env "~/.env" "CUSTOM_VAR" "custom_value"))

;; Load from multiple .env files (setq agent-shell-anthropic-claude-environment (agent-shell-make-environment-variables :load-env '("~/.env" ".env.local") :inherit-env t)) #+end_src

The .env files should contain variables in the format KEY=value, with one variable per line. Comments (lines starting with #) and empty lines are ignored.

*** Anthropic Claude

For login-based authentication (default):

#+begin_src emacs-lisp (setq agent-shell-anthropic-authentication (agent-shell-anthropic-make-authentication :login t)) #+end_src

For API key authentication:

#+begin_src emacs-lisp ;; With string (setq agent-shell-anthropic-authentication (agent-shell-anthropic-make-authentication :api-key "your-anthropic-api-key-here"))

;; With function (setq agent-shell-anthropic-authentication (agent-shell-anthropic-make-authentication :api-key (lambda () (auth-source-pass-get 'secret "anthropic-api-key")))) #+end_src

For OAuth token authentication (the =CLAUDE_CODE_OAUTH_TOKEN= we get from =claude setup-token=):

#+begin_src emacs-lisp ;; With string (setq agent-shell-anthropic-authentication (agent-shell-anthropic-make-authentication :oauth "your-oauth-token-here"))

;; With function (setq agent-shell-anthropic-authentication (agent-shell-anthropic-make-authentication :oauth (lambda () (auth-source-pass-get "secret" "anthropic-oauth-token")))) #+end_src

For alternative Anthropic-compatible API endpoints, configure via environment variables:

#+begin_src emacs-lisp (setq agent-shell-anthropic-claude-environment (agent-shell-make-environment-variables "ANTHROPIC_BASE_URL" "https://api.moonshot.cn/anthropic" "ANTHROPIC_MODEL" "kimi-k2-turbo-preview" "ANTHROPIC_SMALL_FAST_MODEL" "kimi-k2-turbo-preview")) #+end_src

*** Google Gemini

For login-based authentication (default):

#+begin_src emacs-lisp (setq agent-shell-google-authentication (agent-shell-google-make-authentication :login t)) #+end_src

For API key authentication:

#+begin_src emacs-lisp ;; With string (setq agent-shell-google-authentication (agent-shell-google-make-authentication :api-key "your-google-api-key-here"))

;; With function (setq agent-shell-google-authentication (agent-shell-google-make-authentication :api-key (lambda () (auth-source-pass-get 'secret "google-api-key")))) #+end_src

For Vertex AI authentication:

#+begin_src emacs-lisp (setq agent-shell-google-authentication (agent-shell-google-make-authentication :vertex-ai t)) #+end_src

*** OpenAI Codex

For login-based authentication (default):

#+begin_src emacs-lisp (setq agent-shell-openai-authentication (agent-shell-openai-make-authentication :login t)) #+end_src

For API key authentication:

#+begin_src emacs-lisp ;; With string (setq agent-shell-openai-authentication (agent-shell-openai-make-authentication :api-key "your-openai-api-key-here"))

;; With function (setq agent-shell-openai-authentication (agent-shell-openai-make-authentication :api-key (lambda () (auth-source-pass-get 'secret "openai-api-key")))) #+end_src

*** Goose

By default, authentication is handled by Goose's built-in configuration. This needs no =agent-shell= configuration.

Optionally, pass an OpenAI API key from =agent-shell= instead:

#+begin_src emacs-lisp ;; With string (setq agent-shell-goose-authentication (agent-shell-make-goose-authentication :openai-api-key "your-openai-api-key-here"))

;; With function (setq agent-shell-goose-authentication (agent-shell-make-goose-authentication :openai-api-key (lambda () (auth-source-pass-get 'secret "openai-api-key")))) #+end_src

*** Qwen Code

For OAuth login-based authentication:

#+begin_src emacs-lisp (setq agent-shell-qwen-authentication (agent-shell-qwen-make-authentication :login t)) #+end_src

For an OpenAI-compatible provider (for example OpenRouter), pass the API key and set the base URL and model via =agent-shell-qwen-environment=:

#+begin_src emacs-lisp ;; With string (setq agent-shell-qwen-authentication (agent-shell-qwen-make-authentication :openai-api-key "your-openai-api-key-here"))

;; With function (setq agent-shell-qwen-authentication (agent-shell-qwen-make-authentication :openai-api-key (lambda () (auth-source-pass-get 'secret "openrouter-api-key"))))

(setq agent-shell-qwen-environment (agent-shell-make-environment-variables "OPENAI_BASE_URL" "https://openrouter.ai/api/v1" "OPENAI_MODEL" "x-ai/grok-4.3")) #+end_src

*** Cursor

By default authentication is handled externally: run =agent login= once outside Emacs and agent-shell uses that existing login (it sends no ACP authenticate request). This needs no configuration.

Optionally, have agent-shell drive login-based authentication (=cursor_login=):

#+begin_src emacs-lisp (setq agent-shell-cursor-authentication (agent-shell-cursor-make-authentication :login t)) #+end_src

For API key authentication (passed via the =CURSOR_API_KEY= environment variable):

#+begin_src emacs-lisp ;; With string (setq agent-shell-cursor-authentication (agent-shell-cursor-make-authentication :api-key "your-cursor-api-key-here"))

;; With function (setq agent-shell-cursor-authentication (agent-shell-cursor-make-authentication :api-key (lambda () (auth-source-pass-get 'secret "cursor-api-key")))) #+end_src

For auth token authentication (passed via the =CURSOR_AUTH_TOKEN= environment variable):

#+begin_src emacs-lisp (setq agent-shell-cursor-authentication (agent-shell-cursor-make-authentication :auth-token "your-cursor-auth-token")) #+end_src

For no authentication (already authenticated externally):

#+begin_src emacs-lisp (setq agent-shell-cursor-authentication (agent-shell-cursor-make-authentication :none t)) #+end_src

*** Auggie

For login-based authentication (default):

#+begin_src emacs-lisp (setq agent-shell-auggie-authentication (agent-shell-make-auggie-authentication :login t)) #+end_src

For no authentication (when using alternative authentication methods):

#+begin_src emacs-lisp (setq agent-shell-auggie-authentication (agent-shell-make-auggie-authentication :none t)) #+end_src

*** Mistral Vibe

For API key authentication:

#+begin_src emacs-lisp ;; With string (setq agent-shell-mistral-authentication (agent-shell-mistral-make-authentication :api-key "your-mistral-api-key-here"))

;; With function (reusing the API key configured in vibe) (setq agent-shell-mistral-authentication (agent-shell-mistral-make-authentication :api-key (lambda () (string-trim (shell-command-to-string "source ~/.vibe/.env; echo $MISTRAL_API_KEY"))))) #+end_src

*** Opencode + local ollama Qwen2.5

If ~~opencode~~ is not found by ~~agent-shell~~, you can add it to the search path

#+begin_src emacs-lisp (use-package agent-shell :ensure t :config (add-to-list 'exec-path "~/.opencode/bin/")) #+end_src

To register the local Qwen2.5:7b model in opencode, two files needs to be edited:

~~/.config/opencode/opencode.json~~ #+BEGIN_SRC js { "$schema": "https://opencode.ai/config.json", "permission": { "edit": "ask", "bash": "ask", "webfetch": "allow" }, "provider": { "ollama": { "name": "qwen2.5:7b", "npm": "@ai-sdk/openai-compatible", "options": { "baseURL": "http://localhost:11434/v1" }, "models": { "qwen2.5:7b": { "name": "Qwen2.5 7b" } } } } } #+END_SRC

and a dummy auth file ~~/.local/share/opencode/auth.json~~

#+begin_src js { "ollama": { "type": "api", "key": "ollama" } } #+end_src

Huge thanks for /gdindi/ for finding and and describing this in https://github.com/xenodium/agent-shell/issues/526

*** Customizing Available Agents

By default, =agent-shell= includes configurations for all supported agents (Claude Agent, Gemini CLI, Codex, Goose, Grok Build, Cursor, CodeBuddy, Qwen Code, and Auggie). You can customize which agents are available through the =agent-shell-agent-configs= variable.

** Usage

*** Quick Start

=M-x agent-shell= - Start or reuse any of the known agents.

You can select and start any of the known agent shells (see =agent-shell-agent-configs=) via the =agent-shell= interactive command and enables reusing existing shells when available. With a prefix argument (=C-u M-x agent-shell=), it forces starting a new shell session, thus instantiating multiple agent shells.

*** Specific Agent Commands

Start a specific agent shell session directly:

=M-x agent-shell-anthropic-start-claude-code= - Start a Claude Agent session
=M-x agent-shell-auggie-start-agent= - Start an Auggie agent session
=M-x agent-shell-codebuddy-start-agent= - Start a CodeBuddy agent session
=M-x agent-shell-openai-start-codex= - Start a Codex agent session
=M-x agent-shell-google-start-gemini= - Start a Gemini agent session
=M-x agent-shell-goose-start-agent= - Start a Goose agent session
=M-x agent-shell-xai-start-grok= - Start a Grok Build (xAI) agent session
=M-x agent-shell-cursor-start-agent= - Start a Cursor agent session
=M-x agent-shell-kiro-start-agent= - Start a Kiro CLI agent session
=M-x agent-shell-mistral-start-vibe= - Start a Mistral Vibe agent session
=M-x agent-shell-qwen-start= - Start a Qwen Code agent session
=M-x agent-shell-droid-start-agent= - Start a Factory Droid agent session
=M-x agent-shell-pi-start-agent= - Start a Pi coding agent session

*** Setting a default agent for all new shells

You can set a default agent to use for all new shells started via =agent-shell= like so:

#+begin_src emacs-lisp (setq agent-shell-preferred-agent-config (agent-shell-anthropic-make-claude-code-config)) #+end_src

Or by identifier (e.g. Grok Build):

#+begin_src emacs-lisp (setq agent-shell-preferred-agent-config 'grok-build) #+end_src

*** Configuring MCP servers

You can configure MCP servers directly via =agent-shell=. This allows you to avoid having to repeat configurations across every agent that you use.

#+begin_src emacs-lisp (setq agent-shell-mcp-servers '(((name . "notion") (type . "http") (headers . []) (url . "https://mcp.notion.com/mcp")))) #+end_src

** Running agents in Devcontainers / Docker containers (Experimental)

=agent-shell= provides rudimentary support for running agents and shell commands in containers.

Use =agent-shell-command-prefix= to prefix the command that starts the agent, or a shell command that should be run so it is executed inside the container.

*** Static command list

#+begin_src emacs-lisp (setq agent-shell-command-prefix '("devcontainer" "exec" "--workspace-folder" ".")) #+end_src

*** Function-based configuration

For dynamic per-agent containers, provide a function that takes the current agent-shell buffer and returns the command list:

#+begin_src emacs-lisp (setq agent-shell-command-prefix (lambda (buffer) (let ((config (agent-shell-get-config buffer))) (pcase (map-elt config :identifier) ('claude-code '("docker" "exec" "claude-dev" "--")) ('gemini-cli '("docker" "exec" "gemini-dev" "--")) (_ '("devcontainer" "exec" ".")))))) #+end_src

*** Per-session containers

You can use different containers for different shell sessions, even of the same agent type:

#+begin_src emacs-lisp (setq agent-shell-command-prefix (lambda (buffer) ;; Different container based on project (if (string-match "project-a" (buffer-name buffer)) '("docker" "exec" "project-a-dev" "--") '("docker" "exec" "project-b-dev" "--")))) #+end_src

Note that any =:environment-variables= you may have passed to =acp-make-client= will not apply to the agent process running inside the container. It's expected to inject environment variables by means of your devcontainer configuration / Dockerfile.

Next, set an =agent-shell-path-resolver-function= that resolves container paths in the local working directory, and vice versa. Agent shell provides the =agent-shell-devcontainer-resolve-path= function for use with devcontainers specifically: it reads the =workspaceFolder= specified in =.devcontainer/devcontainer.json=, or uses the default value of =/workspaces/= otherwise.

#+begin_src emacs-lisp (setq agent-shell-path-resolver-function #'agent-shell-devcontainer-resolve-path) #+end_src

Note that this allows the agent to access files on your local file-system. While care has been taken to restrict access to files in the local working directory, it's probably possible for a malicious agent to circumvent this restriction.

Optional: to prevent the agent running inside the container to access your local file-system altogether and to have it read/modify files inside the container directly, in addition to setting the resolver function, disable the "read/write text file" client capabilities:

#+begin_src emacs-lisp (setq agent-shell-text-file-capabilities nil) #+end_src

*** Data storage location

By default, agent-shell stores per-project data (transcripts, screenshots, etc.) under a =.agent-shell/= directory in the project root, and automatically adds that directory to =.gitignore= the first time it is created. This is a one-time operation: removing the entry from =.gitignore= will not cause it to be re-added. If you configure =agent-shell-dot-subdir-function= to store data outside the project =.agent-shell/= directory, agent-shell will not update the project's =.gitignore=.

You can change where this data is stored by setting =agent-shell-dot-subdir-function= to a custom function. The function receives one argument, SUBDIR (e.g. ="screenshots"=), and must return the absolute path to that subdirectory (without creating it).

For example, to store data under =user-emacs-directory= instead of the project tree, flattening the project path into a directory name:

#+begin_src emacs-lisp (defun my/agent-shell-dot-subdir (subdir) (let* ((cwd (string-remove-suffix "/" (agent-shell-cwd))) (sanitized (replace-regexp-in-string "/" "-" (string-remove-prefix "/" cwd)))) (expand-file-name subdir (locate-user-emacs-file (concat "agent-shell/" sanitized)))))

(setopt agent-shell-dot-subdir-function #'my/agent-shell-dot-subdir) #+end_src

This stores data at a path like =~/.emacs.d/agent-shell/home-user-src-myproject/screenshots/=.

*** Screenshots from clipboard

You can send a screenshot from your clipboard to your shell with =agent-shell-send-clipboard-image=. Call with =C-u= to =agent-shell-send-clipboard-image-to= to select from your shells. agent-shell relies on external programs to write an image from clipboard to file as configured in =agent-shell-clipboard-image-handlers=. Preconfigured handlers are

=wl-paste= for Wayland desktops,
=xclip= for Xorg,
=pngpaste= for MacOS.
=powershell= for Windows.

*** Inhibiting minor modes during file writes

Some minor modes (for example, =aggressive-indent-mode=) can interfere with an agent's edits. Agent Shell can temporarily disable selected per-buffer minor modes while applying edits.

#+begin_src emacs-lisp (setopt agent-shell-write-inhibit-minor-modes '(aggressive-indent-mode)) #+end_src

*** Keeping the system awake during long turns

Long agent turns can outlast the system idle-sleep timeout, suspending the machine (and the agent) before the turn completes. On Emacs 31.1 or later, =agent-shell-inhibit-system-sleep= (enabled by default) keeps the system awake for the duration of each turn and releases the block once the turn finishes, so the system is only kept awake while there is work in progress (the display is still allowed to blank). This relies on the built-in [[https://git.savannah.gnu.org/cgit/emacs.git/tree/lisp/system-sleep.el][=system-sleep=]] library and has no effect on earlier Emacs versions. It does not prevent "hard" sleep such as closing a laptop lid.

To turn it off:

#+begin_src emacs-lisp (setopt agent-shell-inhibit-system-sleep nil) #+end_src

All of the above settings can be applied on a per-project basis using [[https://www.gnu.org/software/emacs/manual/html_node/emacs/Directory-Variables.html][directory-local variables]].

** Key bindings

=RET= - Submits prompt.
=M-J= - Iserts newline.
=C-c C-c= - Interrupt current agent operation
=TAB and Shift-TAB= - Navigate interactive elements

To customize =RET= binding behaviour, you can use something like:

#+begin_src emacs-lisp :lexical no (use-package agent-shell :bind (:map agent-shell-mode-map ("RET" . newline) ("M-RET" . shell-maker-submit))) #+end_src

Or if you prefer C-c C-c to send and C-c C-k to interrupt, use the following:

#+begin_src emacs-lisp :lexical no (use-package agent-shell :bind (:map agent-shell-mode-map ("RET" . newline) ("C-c C-c" . shell-maker-submit) ("C-c C-k" . agent-shell-interrupt))) #+end_src

*** Session controls

When the active agent advertises the corresponding session option, you can switch it from the shell buffer:

=C-c C-v= - Set model.
=C-c C-m= - Set session mode.
=C-c C-t= - Set thought level (reasoning effort).
=C-= - Cycle session mode.

The current selections are shown in the header right after the agent name; click any segment to open a popup menu. The thought level segment only appears for models that expose it (e.g. Claude Sonnet/Opus with effort levels; Codex's reasoning_effort).

From the viewport (=agent-shell-viewport-view-mode= / =agent-shell-viewport-edit-mode=), the same controls are bound to:

View mode: =v= (model), =s= (mode), =t= (thought level).
Edit mode: =C-c C-v=, =C-c C-m=, =C-c C-t=.

*** Evil Evil users may want to rebind ~~RET~~ for inserting a new line in =insert mode= and sending the prompt in =normal mode=.

Also, when viewing diffs (before accepting changes) it may be annoying having to enter =insert mode= to send keys (~~y/n/p/q/etc~~). If this is your case, you can make these buffers start in Emacs mode (you can always go to Evil modes if you need to with ~~C-z~~). #+begin_src emacs-lisp (use-package agent-shell :config ;; Evil state-specific RET behavior: insert mode = newline, normal mode = send (evil-define-key 'insert agent-shell-mode-map (kbd "RET") #'newline) (evil-define-key 'normal agent-shell-mode-map (kbd "RET") #'comint-send-input)

;; Configure *agent-shell-diff* buffers to start in Emacs state
(add-hook 'diff-mode-hook
    (lambda ()
      (when (string-match-p "\\*agent-shell-diff\\*" (buffer-name))
	(evil-emacs-state)))))

#+end_src

Customizations

#+BEGIN_SRC emacs-lisp :results table :colnames '("Custom variable" "Description") :exports results (let ((rows)) (mapatoms (lambda (symbol) (when (and (string-match "^agent-shell" (symbol-name symbol)) (custom-variable-p symbol)) (push `(,symbol ,(car (split-string (or (documentation-property symbol 'variable-documentation) (get (indirect-variable symbol) 'variable-documentation) (get symbol 'variable-documentation) "") "\n"))) rows)))) (sort rows (lambda (a b) (string< (symbol-name (car a)) (symbol-name (car b)))))) #+END_SRC

#+RESULTS: | Custom variable |--------------------------------- | agent-shell-agent-configs | agent-shell-anthropic-authentication | agent-shell-anthropic-claude-acp-command | agent-shell-anthropic-claude-command | agent-shell-anthropic-claude-environment | agent-shell-anthropic-default-model-id | agent-shell-anthropic-default-session-mode-id | agent-shell-auggie-acp-command | agent-shell-auggie-authentication | agent-shell-auggie-environment | agent-shell-buffer-name-format | agent-shell-busy-indicator-frames | agent-shell-clipboard-image-handlers | agent-shell-completion-mode-hook | agent-shell-command-prefix | agent-shell-context-sources | agent-shell-cursor-acp-command | agent-shell-cursor-environment | agent-shell-display-action | agent-shell-droid-acp-command | agent-shell-droid-authentication | agent-shell-droid-environment | agent-shell-embed-file-size-limit | agent-shell-file-completion-enabled | agent-shell-github-acp-command | agent-shell-github-default-model-id | agent-shell-github-default-session-mode-id | agent-shell-github-environment | agent-shell-google-authentication | agent-shell-google-gemini-acp-command | agent-shell-google-gemini-environment | agent-shell-goose-acp-command | Description | --------------+---------------------------------------------------------------------------------| | The list of known agent configurations. | | Configuration for Anthropic authentication. | | Command and parameters for the Anthropic Claude client. | | Command and parameters for the Anthropic Claude client. | | Environment variables for the Anthropic Claude client. | | Default Anthropic model ID. | | Default Anthropic session mode ID. | | Command and parameters for the Auggie client. | | Configuration for Auggie authentication. | | Environment variables for the Auggie client. | | Format to use when generating agent shell buffer names. | | Frames for the busy indicator animation. | | Handlers for saving clipboard images to a file. | | Hook run after entering or leaving ‘agent-shell-completion-mode’. | | Command prefix for executing commands in a container. | | Sources to consider when determining M-x agent-shell automatic context. | | Command and parameters for the Cursor agent client. | | Environment variables for the Cursor agent client. | | Display action for agent shell buffers. | | Command and parameters for the Factory Droid ACP client. | | Configuration for Factory Droid authentication. | | Environment variables for the Factory Droid ACP client. | | Maximum file size in bytes for embedding with ContentBlock::Resource. | | Non-nil automatically enables file completion when starting shells. | | Command and parameters for the GitHub Copilot agent client. | | Default GitHub Copilot model ID. | | Default GitHub Copilot session mode ID. | | Environment variables for the GitHub Copilot agent client. | | Configuration for Google authentication. | | Command and parameters for the Gemini client. | | Environment variables for the Google Gemini client. | | Command and parameters f

Agent Shell for Emacs

At a glance

README

Comments (0)