What is Dagu?

Dagu is a local-first workflow engine. It is declarative, file-based, self-contained, and air-gapped ready. It runs as a single binary and scales from laptop to distributed cluster.

Key Capabilities

• Declarative YAML - JSON Schema validation with clear error messages
• Custom step types - Reusable wrappers around builtin step types with JSON Schema-validated input
• Composable workflows - Nest sub-DAGs with parameters (depth limited only by available memory)
• Distributed execution - Route tasks to workers via labels (GPU, region, etc.)
• Scheduler - Cron expressions with start/stop/restart support
• AI Agent integration - Built-in AI agent for workflow management and workflow steps operate on your computer
• Durable execution - Automatic retries, hooks, and state persistence for long-running workflows
• Web UI - Monitor, control, and debug workflows in real-time

The Dagu Difference: Keep workflow orchestration separate from business logic. Define workflows declaratively, stay zero-invasive to application code, and get a more capable alternative to cron without taking on Airflow-level complexity.

AI Agent

Dagu includes a built-in LLM-powered agent that can read, create, and modify your workflows. Use it interactively through the Web UI chat, or add type: agent steps to your DAGs for automation.

• Create and update workflows — describe what you need; the agent generates valid DAG YAML, validates it against the schema, and opens it in the UI
• Debug and fix failures — point the agent at a failed run; it reads logs, suggests probable causes, and can patch configurations
• Answer questions — ask what a DAG does, how to configure an executor, or why a step failed

The agent runs as an LLM tool-calling loop with configurable tools (shell execution, file read/write, schema lookup, sub-agent delegation) and safety controls (RBAC enforcement, per-tool enable/disable, regex-based bash command policies). Results are non-deterministic and should be reviewed.

Bring your own model: Anthropic, OpenAI, Gemini, OpenRouter, and local servers (Ollama, vLLM) are supported. Models and API keys are configured in the Web UI at /agent-settings.

steps:
  - id: analyze_logs
    type: agent
    messages:
      - role: user
        content: |
          Analyze the error logs at /var/log/app/errors.log from the last hour.
          Summarize the root causes and suggest fixes.
    output: ANALYSIS_RESULT

See Agent Overview and Agent Step for full documentation.

Workflow Operator

Workflow Operator is Dagu's persistent AI operator for Slack and Telegram. It uses a Slack or Telegram bot to bridge your messaging platform with the built-in AI agent, so you can manage workflows without opening the Web UI.

• Debug issues - ask the agent to check logs and diagnose failures
• Recover from incidents - re-run workflows with adjusted parameters through chat
• Get notified - receive DAG run completion notifications with AI-generated summaries
• Approve actions - respond to approval gates via interactive buttons in Slack or Telegram

Each conversation maps to a persistent agent session. The Slack or Telegram bot connector supports safe mode with configurable bash command policies.

# Enable Workflow Operator on Telegram
bots:
  provider: telegram
  safe_mode: true
  telegram:
    token: ${TELEGRAM_BOT_TOKEN}
    allowed_chat_ids:
      - 123456789

See Workflow Operator for setup instructions.

How It Works

Dagu executes workflows defined as steps in YAML. Steps form a Directed Acyclic Graph (DAG), ensuring predictable execution order.

Sequential Execution

type: chain
steps:
  - command: echo "Step 1"
  - command: echo "Step 2"

Parallel Execution

type: graph
steps:
  - id: step_1
    command: echo "Step 1"
  - id: step_2a
    command: echo "Step 2a"
    depends: [step_1]
  - id: step_2b
    command: echo "Step 2b"
    depends: [step_1]
  - id: step_3
    command: echo "Step 3"
    depends: [step_2a, step_2b]

Architecture Overview

Local mode: CLI, Web UI, or API triggers the Agent, which runs steps via Executors.

Distributed mode: Scheduler dispatches work to a Coordinator, which routes tasks to Workers based on label selectors (e.g., gpu=true, region=us-east).

See Architecture for details.

Built-in Step Types

Type	Description
command	Shell commands (bash, sh, PowerShell, cmd)
agent	LLM-powered agent with tool-calling loop
ssh	Remote command execution via SSH
sftp	Remote file transfer via SFTP
docker	Container execution with volume mounts and registry auth
k8s, kubernetes	Run a step as a Kubernetes Job with DAG-level defaults
harness	Run CLI-based coding agents and custom harness adapters
http	HTTP/REST API requests
jq	JSON query and transformation
template	Text rendering with Go templates and structured data
mail	Email notifications with attachments
dag	Sub-workflow execution (hierarchical composition)
router	Route execution by evaluated value
postgres, sqlite	SQL queries
redis	Redis commands and scripts
s3	S3 object operations
chat	LLM chat completion

approval is not a step type. It is a field available on steps to pause a workflow for human review. See Approval.

You can also define reusable step_types in a DAG or base config. See Custom Step Types and Step Types Reference for the exact configuration surface.

Learn More

• Quick Start - Running in minutes
• Core Concepts - Workflows, steps, and dependencies
• AI Agent - Built-in LLM agent for workflow management
• Workflow Operator - Manage workflows from Slack or Telegram
• Architecture - System internals and distributed execution
• Examples - Ready-to-use workflow patterns