Building Workflows

This guide walks you through building workflows in MIFY — from simple chains to complex multi-step pipelines with branching, human approval, and AI orchestration.

MIFY Canvas — visual workflow editor

Workflow Basics

A workflow in MIFY is a directed acyclic graph (DAG) — a series of connected steps that process data in one direction without loops.

Trigger → Node A → Node B → Node C (output)
                 ↘ Node D ↗

Every workflow has three elements:

Trigger — how the workflow starts (manual, webhook, schedule, event, or form submission)
Nodes — processing steps (AI calls, data transforms, logic gates, integrations)
Edges — connections between nodes that define data flow

Node Categories

MIFY provides 126 native nodes across 24 categories:

Category	Examples	Purpose
LLM	Chat, Embed, Image Generate, Vision, ASR, TTS	AI model operations
Transform	Template, JavaScript	Data transformation
Logic	Route If, Route Switch, Loop Map, Sub-Workflow	Flow control
RAG	Ingest, Chunk, Embed, Store, Retrieve, Rerank	Retrieval-augmented generation
Platform	Dify, Flowise, Langflow, n8n, and 20+ more	External platform integration
Agent	ReAct, Multi-Agent (Plan-and-Execute, Reflection, Supervisor, Swarm), CLI Agent, Browser Use, Computer Use	AI agent orchestration
Safety	PII Detector, Injection Detector, Jailbreak Detector	Content protection
Enterprise	Email, Calendar, Notifications, Quota, Policy	Business integrations
Cloudflare	D1, KV, R2, Vectorize, Queues	Edge storage and compute
Sandbox	sandbox.exec, sandbox.plan, sandbox.session, host.apply, workspace	Isolated code execution

Step-by-Step: Build a RAG Workflow

Let’s build a workflow that answers questions using your own documents.

1. Create the Trigger

Drag a Manual Trigger onto the canvas.

2. Add Document Retrieval

Drag a Retrieve node. Configure it:

Knowledge Base: Select your knowledge base (create one at /knowledge first)
Top K: 5 (return top 5 relevant chunks)

Connect the trigger to the Retrieve node.

3. Add the AI Response

Drag a Chat node. Configure it:

System Prompt: “Answer the user’s question based on the following context: {{context}}”
Provider: Your preferred AI provider
Model: Your preferred model

Connect the Retrieve node to the Chat node. The retrieved documents flow as context.

4. Run and Test

Click Run, enter a question, and see the AI respond using your documents as context.

Templates

MIFY includes 149 built-in templates for common patterns:

Category	Templates	Examples
RAG	9	Document Q&A, Knowledge Base Chat
LLM	12	Text Summary, Translation, Classification
Logic	17	Conditional Routing, Batch Processing
Platform Demos (Track A + B)	60+	Flowise Import, Dify Bridge, n8n Connector, Strapi, Directus, Supabase, WordPress, Saleor, MinIO, Mautic, Authentik, Keycloak
Agent	5	Multi-Agent strategies, Research Agent

Browse templates at /templates in the UI. Click any template to preview it, then Use Template to add it to your canvas.

Human-in-the-Loop (HITL)

Add human approval gates to any workflow:

Drag a Pause node between any two nodes
When execution reaches the Pause node, it stops and waits
A notification appears in your Inbox (/inbox)
Review the data, then Approve or Reject
Approved: execution continues. Rejected: execution stops.

Use HITL for:

Reviewing AI-generated content before publishing
Approving financial transactions
Validating data extraction results
Any decision that needs human judgment

Multi-Agent Strategies

The Multi-Agent node (mify-agent-multi_agent) coordinates several agents inside a single execution. Pick a strategy when you configure the node:

Strategy	What it does	When to use
Plan-and-Execute	Decomposes the goal into subtasks, then executes each sequentially	Multi-step research or planning tasks
Reflection	Generates an answer, then critiques and revises it in a loop	Quality-sensitive output (writing, code review)
Supervisor	One coordinator routes work to specialized worker agents	Tasks that need different specialist roles
Swarm	Peer-to-peer agent network with a shared message bus	Open-ended exploration, parallel research

See the Multi-Agent node reference for configuration.

Versioning Your Workflows

Every published workflow has a version history at /graphs/[graphId]/versions:

Publish — capture the current draft as a GraphVersion snapshot (idempotent)
Archive — archive the currently-published version
Diff — line-level diff between any two versions (uses diff-match-patch)
Draft from version — fork a prior version into a new draft to resurrect/edit it

Each transition is logged for audit. See Execution & Trust.

Form-Triggered Workflows

Trigger any workflow from a multi-page form submission. Three pages back the feature:

/forms — list your forms
/forms/[formId]/builder — drag-drop builder with conditional fields and multi-page navigation
/forms/[formId] — public render of the submittable form

Submissions hit POST /api/forms/:id/submit, which calls the FormTriggerPlugin to start the workflow. See Franchise Operators — Embedded Widgets for the embed flow.

Realtime Collaboration

Multiple users can edit a workflow at the same time. The canvas runs on Hocuspocus + Yjs:

Each session opens a WebSocket to the Hocuspocus server
Auth handshake uses SuperTokens — only authorized users can connect
Live presence cursors show who else is in the workflow (with name + color)
Edits sync in real time and are periodically saved back to the canonical MifyGraph so versioning continues to work
Y.Doc binary state is persisted in Postgres (GraphYjsState table) — connections can reattach without losing work

Durable Execution + Time-Travel

Switch a workflow to durable mode (in /graphs/[graphId]/settings) to run it on Temporal:

Each node runs as a Temporal Activity — durable, replayable, horizontally scalable
Long-running flows survive process restarts without losing progress
Full event history captured for every run

Use the time-travel UI at /workflows/[graphId]/runs/[runId] to:

Replay from any prior step
Fork with overridden inputs to explore alternatives
Inspect intermediate state at any point

The /admin/temporal/durable admin page lists all durable runs and surfaces failures for replay.

Workflow as API

Publish a workflow, get a stable HTTP endpoint with auto-generated OpenAPI. From /graphs/[graphId]/versions, the WorkflowEndpointInfo panel shows you:

The endpoint URL (POST /api/workflow-endpoints/:slug)
A sample curl
Per-API-key rate-limit policy
A link to the auto-generated OpenAPI 3.1 spec for that workflow

The same Published workflow is also exposed as:

An MCP tool at /api/workflow-mcp/* — any MCP client (Claude Desktop, Cursor, etc.) can list and call it
An A2A v1.0 agent at /api/a2a-server/* — with auto-populated AgentCard.skills
A LangGraph Agent Protocol agent at /api/agent-protocol/* — bidirectional bridge

Execution Modes

Mode	Use Case	Behavior
Experimental	Development	Minimal overhead, fast iteration
Audited	Compliance	Full audit trail, integrity hashing

Importing from Other Platforms

Already using another workflow platform? Import directly:

Platform	How
Flowise	`/canvas` → Import → Paste Flowise JSON
Dify	`/canvas` → Import → Paste Dify YAML
Langflow	`/canvas` → Import → Paste Langflow JSON
n8n	`/canvas` → Import → Paste n8n JSON

MIFY maps foreign nodes to native equivalents where possible. Unmapped nodes become “opaque” nodes that delegate to the original platform.

Content Generation

MIFY includes a dedicated content generation system at /content-gen:

Path	How It Works	Best For
MIFY Queue	Submits ComfyUI workflows to GPU providers	High-quality image/video
AI Provider	Direct calls to AI providers (sync or streaming)	Text, quick image gen
ComfyUI Local	Runs workflows on your local ComfyUI	Custom models, local dev

Next Steps

API Reference — Programmatic workflow management
Troubleshooting — Common issues
Overview — Full platform capabilities