AI Agent Panel

Skim integrates with AI agents in two ways:

1. Agent panel: a built-in chat interface (Ctrl-e) that speaks the Agent Client Protocol (ACP).
2. External integration: CLI commands (skim session …) and an MCP server so agents can drive a running Skim instance.

The agent panel

Press Ctrl-e to toggle a chat panel alongside your diff. Agents are spawned as subprocesses that communicate over stdio. There is no daemon.

# 1. Configure agents in ~/.skim/config.json (see below)
# 2. Open your diff
skim --staged
# 3. Press Ctrl-e to toggle the panel

Configuration

Configure agents and panel settings in ~/.skim/config.json:

{
  "agent_servers": {
    "Claude Code": {
      "command": "claude-code-acp",
      "skim": { "default": true, "model": "opus" }
    },
    "Codex": {
      "command": "codex",
      "protocol": "codex",
      "approval_policy": "never",
      "sandbox_mode": "workspace-write",
      "web_search": true
    },
    "OpenCode": {
      "command": "opencode",
      "args": ["acp"]
    }
  }
}

Field	Type	Required	Description
command	string	Yes	CLI command to spawn the agent
args	string[]	No	Extra CLI arguments
env	object	No	Environment variables (${VAR} expansion supported)
protocol	string	No	"acp", "opencode", or "codex"
approval_policy	string	No	Codex approval mode ("never", "on-request")
sandbox_mode	string	No	Codex sandbox ("read-only", "workspace-write", "danger-full-access")
web_search	bool	No	Enable Codex’s native web-search tool
skim.default	bool	No	Auto-connect to this agent (default false)
skim.model	string	No	Model to use (e.g. "opus", "sonnet")
skim.mode	string	No	Session mode (e.g. "plan", "code")

For native Codex, set "protocol": "codex" with "command": "codex". Skim launches codex app-server and applies sandbox_mode / web_search as first-class settings.

Agent selection:

• Single agent: auto-connects immediately.
• Multiple agents with a default: auto-connects to the one marked "default": true.
• Multiple agents, no default: shows a selection menu (j / k to navigate, Enter to select).

Switch agents anytime via the command palette (Ctrl-p, then >Switch Agent).

Option	Values	Default
agent_panel_side	"left", "right"	"right"

@file references

Type @ in the prompt to fuzzy-search and embed file contents:

@src/m     → matches src/main.zig, src/modes/*, etc.
@readme    → matches README.md

Key	Action
@	Open the file picker (at a word boundary)
↑ / ↓ or Ctrl-p / Ctrl-n	Navigate the file list
Enter or Tab	Insert the selected file
Ctrl-C / ESC	Close the file picker

Agent-panel keybindings

The panel uses vim-style modal editing.

Global

Key	Action
Ctrl-E	Close the panel, return to the diff
Ctrl-G	Edit the prompt in $EDITOR
Ctrl-S	Stash / unstash the prompt
Ctrl-T	Toggle the todo-list expansion

Insert mode

Key	Action
Enter	Send the prompt to the agent
Ctrl-J	Insert a newline
Ctrl-C / ESC	Exit to normal mode
/	Show the slash-command menu (at prompt start)
@	Show the file picker (at a word boundary)
!	Toggle shell-command mode (empty input)
Up	Restore the stashed prompt (empty input)

Normal mode

Key	Action
Ctrl-W h/j/k/l	Focus a pane / diff edge
Ctrl-W w	Cycle panes / diff focus
Ctrl-W v / Ctrl-W s	Open a vertical / horizontal split
Ctrl-W c / Ctrl-W o	Close a pane / keep only this pane
Ctrl-W H/J/K/L	Move the pane to an edge
Ctrl-W + - < >	Resize the focused pane
i / a / I / A	Enter insert mode
h / l	Move the cursor left / right
w / b / e	Word motions
0 / $	Line start / end
gg / G	Jump to top / bottom of the input
Ctrl-D / Ctrl-U	Half-page down / up in the input
x / dd	Delete char / line
:	Open the command palette
?	Show help
gb / Space+b	Enter history mode
gt / gT	Next / previous tab
Space+f	Scroll to bottom, enable follow
V	Toggle the diff view mode
Tab	Cycle session modes
Ctrl-C	Interrupt the agent

History mode

Enter with gb or Space+b to browse and yank from message history.

Key	Action
j / k	Move the cursor down / up
h / l	Jump to the previous / next message
gg / G	Jump to top / bottom
Ctrl-D / Ctrl-U	Page down / up
M	Center the cursor in the viewport
v	Enter visual selection mode
y	Yank the user message at the cursor
yy	Yank the current line
Y	Yank the entire current message
Space+f	Resume follow mode, exit history
i	Exit to insert mode
Ctrl-C / ESC / q	Exit to normal mode

Visual mode

Enter with v from history mode.

Key	Action
j / k	Extend the selection down / up
y	Yank the selection to the clipboard
Ctrl-C / ESC / v	Exit visual mode

Permission prompt

When the agent requests permission:

Key	Action
j / k or Up / Down	Navigate options
Ctrl-D / Ctrl-U	Scroll message history
Enter / y	Accept the selected option
Ctrl-C / ESC / n	Reject / cancel

Menus (slash, file picker, command palette)

Key	Action
Ctrl-N / Ctrl-P	Navigate the menu
Tab	Insert the selected item
Enter	Insert and execute
Ctrl-C / ESC	Close the menu

Slash commands

Command	Action
/clear	Clear the session and start fresh
/model	Switch the AI model
/resume	Resume a previous session

MCP server

For agents that speak the Model Context Protocol, add Skim to the agent’s config:

{
  "mcpServers": {
    "skim": {
      "command": "skim",
      "args": ["mcp", "--stdio"]
    }
  }
}

Available MCP tools:

Tool	Description
list_clients	List all connected Skim TUI instances
get_diff_context	Get diff metadata (files, stats, mode)
get_file_diff	Get the full diff for a specific file
add_comment	Add a review comment to a specific line
get_comments	Get all comments from a Skim instance

The skim session CLI provides the same capabilities as fallback commands when MCP isn’t available.

Claude Code skill

Install the Skim skill to teach Claude Code how to review with Skim:

npx skills add ctdio/skim

Then ask your agent to review with Skim:

Use skim to review this code, post comments in areas I should focus on

The agent finds Skim sessions running in the same directory and drives them via the MCP tools (with CLI fallbacks).