Dify vs Langflow (2026)

Dify connects to model providers, tools, data sources and external services through a Tool node, an HTTP Request node, MCP tools, custom endpoints and a marketplace of installable plugins, and agent nodes call tools with per-parameter control.

Core components and provider bundles cover models, embeddings, vector stores, loaders and tools, with an API Request component, custom Python components, Tool Mode and an MCP client.

Dify workflows support branching, iteration, loops, classifier routing and sandboxed code, with agent nodes running bounded tool loops inside a flow and workflows callable as tools or MCP servers.

Flows are typed node graphs with loops, branching and reusable grouped components, and flows can serve as tools for an agent.

Workflows start from Schedule, plugin and Webhook triggers as well as on demand, metered as Trigger Events on every tier, and publish to a web app, embed, REST API, CLI or MCP server.

Each flow gets its own authenticated webhook endpoint, alongside the REST API, an embeddable chat component, the Playground and MCP tools; no built-in scheduler.

Knowledge bases are a first-class surface with configurable chunking and indexing, a custom Knowledge Pipeline, ingestion from files, Notion and website crawl, a retrieval node, hit testing and an External Knowledge API.

Retrieval is built from loaders, splitters, embedding models and vector store bundles the customer chooses and owns.

State persists within a conversation through conversation variables, chat history and an agent memory window, but no cross-session or learned memory is documented.

Memory bases store long-term chat history in a vector store across sessions, and chat memory by session ID is on by default, editable and deletable.

A Human Input node pauses a workflow so a person can approve, edit, comment on or forward the run before it continues, with timeout handling, alongside a Sensitive Content Moderation endpoint.

A Human Input component pauses a flow for approve or reject branches, and agents can require approval per tool, resuming from a checkpoint.

Dify states SOC 2 Type I and Type II, ISO 27001:2022 and GDPR with a trust center, and offers SSO, workflow-level RBAC, SCIM, session policies, SIEM-streamed audit logs and bring-your-own-key encryption on Enterprise.

API key, JWT and external OIDC or SSO-gateway authentication with a pluggable role layer, but open-source Langflow does not enforce per-resource roles and no attestation applies.

Dify logs runs with per-node inputs, outputs and errors, a run history and variable inspector, an analytics dashboard, OpenTelemetry trace export to seven named tools, and tamper-evident audit logs on Enterprise.

Each component's output and logs can be inspected after a run, and monitor endpoints and the Playground expose message history per session.

Dify runs as hosted cloud, a single-tenant dedicated cluster in the customer's region, a customer-managed deployment in the customer's own AWS, GCP or Azure VPC, fully air-gapped on-prem, or as the self-hosted open-source Community Edition.

MIT-licensed and self-hosted, as containers, a Python package or a desktop app, on a database the operator controls.

Dify Marketplace offers a filterable catalog of named ready-to-use workflow templates with publishers and usage counts, alongside a typed plugin catalog and a flow for customers to publish their own apps.

Pre-built templates such as Basic Prompting and the Memory Chatbot are ready to use, alongside a Starter Project folder, reusable grouped components and JSON import.

The customer chooses models per workflow node, using Dify credits on hosted OpenAI, Anthropic, Gemini, xAI and Tongyi models or its own provider keys, and air-gapped deployments can run local models.

Provider and model are chosen per node across OpenAI, Anthropic, Google, Cohere, Hugging Face and local Ollama models, with custom components for others.

Dify is callable from outside through a REST API with a published OpenAPI specification, the difyctl CLI, a plugin development kit, inbound webhooks and custom endpoints, and published workflows can be exposed as MCP servers.

Flows are exposed as MCP servers and Langflow connects to external MCP servers, alongside a documented REST API, JSON import and export, custom Python components, an embeddable chat component and LFX.

Dify offers node and flow debugging against sample input, run history, error-handling branches, retrieval hit testing, annotations and version restore, but no evaluation, scoring or regression product.

Strong debugging, with the Playground, single-component runs, inspection, freeze and editable message logs, but no evaluation, scoring or regression capability.

No browser, desktop or computer-use node exists; Dify reaches content and systems over HTTP, APIs, MCP tools and plugins rather than by operating a rendered interface.

No first-party browser or computer-use component; web search is scraping, and browser automation arrives only through an Apify bundle or community components.