Agent

The Dagu agent is an LLM-powered assistant integrated into the Web UI. It can read, create, and modify your workflows through a chat interface with tool-calling capabilities.

Accessing the Agent

Click the Agent button at the bottom-left corner of any page. The button shows a yellow pulsing indicator when the agent is processing.

Setup

All agent configuration is managed through the Web UI at /agent-settings (requires admin role).

1. Enable the Agent

Toggle the agent on from the settings page. Alternatively, set the environment variable DAGU_AGENT_ENABLED=true.

This is the only setting configurable via environment variable. Everything else is configured through the Web UI.

2. Add a Model

Click Add Model to configure an LLM provider. Each model has:

Field	Required	Description
Name	yes	Display name
Provider	yes	anthropic, openai, gemini, openrouter, zai, or local
Model	yes	Model identifier (e.g., claude-sonnet-4-5, gpt-4o)
API Key	yes*	Provider API key. Not required for local provider.
Base URL	no	Custom endpoint. Required for local provider (e.g., http://localhost:11434/v1).
Context Window	no	Token context size
Max Output Tokens	no	Maximum output tokens
Input Cost / 1M	no	Cost per 1M input tokens (for cost tracking)
Output Cost / 1M	no	Cost per 1M output tokens (for cost tracking)
Supports Thinking	no	Enable extended thinking capability

The settings page includes presets for common models (Claude, GPT-4, Gemini, etc.) that pre-fill these fields.

3. Set a Default Model

Click the star icon next to a model to set it as the default. The agent uses this model unless overridden per-session or per-step.

Chat Interface

Sessions

• Each session is persisted per user
• Select previous sessions from the sidebar
• Click + in the header to start a new session
• Each session tracks cost (token usage × model pricing)

DAG Context

The agent receives DAG context automatically:

1. Automatic: When viewing a DAG or run page, that DAG's name, file path, run ID, and status are included
2. Manual: Use the DAG picker in the input area to attach additional DAGs

Model and Soul Selection

The input area includes dropdowns to override the default model and select a soul (personality) for the current message.

Safe Mode

Toggle the shield icon in the header to enable/disable safe mode:

• On (default): Bash commands that match deny rules trigger an approval prompt
• Off: All commands execute immediately without approval

Sending Messages

• Type in the input area at the bottom
• Press Enter to send (Shift+Enter for new line)
• Click Stop (square icon) to cancel processing

Available Tools

Tool	Description
bash	Execute shell commands (120s default timeout, 600s max)
read	Read file contents with line numbers
patch	Create, edit, or delete files
think	Record reasoning without side effects
navigate	Open pages in the Dagu UI
ask_user	Prompt the user with options or free-text input
read_schema	Look up DAG YAML schema documentation
delegate	Spawn sub-agents for parallel tasks
use_skill	Execute a skill from the skill store (when skills are configured)
search_skills	Search available skills by query (when skills are configured)
remote_agent	Delegate tasks to agents on remote nodes (when remote nodes are configured)
list_remote_nodes	List available remote nodes (when remote nodes are configured)

Tools can be individually enabled/disabled in the Tool Policy section of the settings page.

See Tools Reference for full parameter documentation.

Bash Command Policy

Bash command execution is controlled by a configurable policy with regex-based rules. Configure this in the Tool Policy section of the settings page.

Setting	Values	Default	Description
Default Behavior	allow, deny	allow	Action when no rule matches
Deny Behavior	ask_user, block	ask_user	What happens when a command is denied. ask_user shows an approval prompt; block silently rejects.

Rules are evaluated in order. Each rule has:

Field	Description
Pattern	Regex matched against the command
Action	allow or deny
Enabled	Toggle the rule on/off

Example: to require approval for destructive commands while allowing everything else, set default behavior to allow and add deny rules:

Pattern: ^rm\s+      Action: deny
Pattern: ^chmod\s+    Action: deny

When a command is denied with ask_user behavior, a prompt appears showing the command and working directory. The user can approve or reject. Approval times out after 5 minutes.

The policy also blocks shell constructs that could bypass rule matching: backticks, $(...) command substitution, heredocs (<<), and process substitution (<(...), >(...)).

Configuration Storage

Agent configuration is stored in {DAGU_HOME}/data/agent/config.json. The actual structure:

{
  "enabled": true,
  "defaultModelId": "claude-sonnet",
  "toolPolicy": {
    "tools": { "bash": true, "read": true, "patch": true, "...": true },
    "bash": {
      "rules": [],
      "defaultBehavior": "allow",
      "denyBehavior": "ask_user"
    }
  },
  "enabledSkills": [],
  "selectedSoulId": "",
  "webSearch": { "enabled": false }
}

Models are stored separately and managed through their own API. Do not edit these files directly — use the Web UI at /agent-settings.

See Also

• Agent Step — Run the agent as a workflow step (type: agent)
• Tools Reference — Full parameter documentation for each tool
• Memory — Persistent context across sessions
• Souls — Configurable agent personalities
• Chat & AI Agents — type: chat for simpler LLM calls in DAG workflows