Workflow Studio

Design AI Workflows That Hold up in Production.

User interactions should lead to reliable results. With Dify Workflow Studio, teams can design the logic in between on a collaborative visual canvas, combining model calls, knowledge retrieval, tools, code, branching, triggers, and human review to test and ship AI apps as real solutions that work.

Get Started

Harness the Power of Agents.

A model is one node. What makes it a reliable agent is the engineering around it. A workflow is that harness.

Human Judgement In The Loop

Pause before actions that affect sensitive data, access, or policy. Let a person approve, edit, comment, or forward the run before it continues.

Trace Every Run in Detail

Inspect node outputs, variables, execution paths, and logs so all transactions are traced to retrieval, the model, the tool, the user, the device, or any data.

Turn Existing Processes Into Workflows

Translate inputs, retrieval, model calls, tool use, branches, and outputs into connected nodes, so the logic your business follows becomes something you can build.

One Canvas, Multiple Ways to Run.

Workflow and Chatflow share the same node library and execution model. They differ in how runs are triggered and how users interact with the result.

One Pass, End to End.

For automation, batch jobs, document processing, and backend pipelines. A Workflow starts from an input or trigger, runs through the designed path, and returns a result.

One Run Per Message.

For assistants, guided Q&A, and structured support. Each user message travels through the flow before a reply is sent, with conversation memory preserved across turns.

Workflows Can Start Themselves.

Trigger runs without a user in the loop: scheduled jobs, plugin events, and inbound webhooks.

Build Around The Real Handoffs.

Useful AI apps do more than call a model. They classify the request, retrieve context, branch on data, call approved systems, and pause when a human needs to decide.

Nodes we have

User Input
Scheduled trigger
LLM
Knowledge Retrieval
Question Classifier
If / Else
Iteration
Tool
Code
Human Input

Nodes types

Start
Begin from a user message, API call, scheduled trigger, webhook, plugin event, or uploaded payload.
Reason & retrieve
Call models, classify intent, extract fields, and ground responses in prepared knowledge.
Control flow
Branch with conditions, iterate over lists, loop until a result is reached, or merge paths.
Act
Use Dify tools, custom APIs, MCP tools, HTTP requests, or sandboxed code.
Human review
Collect approval, edits, comments, forwarding, or timeout handling before the run continues.

Test, Version, and Watch Every Run.

Prototype is production — the same workflow runs throughout, with debug, reliability, versions, and traces at every step.

Debug

Run a Node, or the Whole Flow.

Test any part of the canvas with sample input, then inspect intermediate values before real users hit the app.

Reliability

Design Predictable Failure Paths.

LLM, HTTP, Code, and Tool nodes can stop, return defaults, or route to recovery branches with typed error metadata.

Versions

Restore Any Published Build.

Keep release history, restore published versions, and move workflows across workspaces with DSL export.

Observability

Send Runs to Your Observability Stack.

Stream traces into Langfuse, LangSmith, Arize, W&B Weave, and more.

Publish Once. Surface Everywhere.

Once the logic is ready, expose it where users and other systems need it: as a hosted app, API endpoint, MCP-compatible tool, or reusable template.

API
Generate app credentials, stream responses, manage sessions, and integrate from your backend.
MCP Server
Expose workflows as native tools for Claude Desktop, Cursor, and other MCP-aware clients.
Template
Export DSL or publish to the Template Marketplace so teams can reuse the design.
Web app
Share a hosted app link for chat, workflow forms, sessions, citations, and batch processing.