Dify

Short answer. Dify is an open-source LLM application development platform by LangGenius that combines AI workflow, RAG pipeline, agent capabilities, model management and observability in one workspace — the leading model-neutral, self-hostable alternative to vendor-bound agent builders.

Key facts.

• Visual workflow canvas plus a RAG pipeline, agent layer and model-management layer, with observability integrations (Opik, Langfuse, Arize Phoenix).
• Model-agnostic: integrates hundreds of proprietary and open-source LLMs, including any OpenAI API-compatible model.
• Three deployment modes: Dify Cloud (managed SaaS), VPC, and self-hosted via Docker Compose (docker compose up -d).
• Open source under the “Dify Open Source License” — Apache-2.0 based with additional conditions; a commercial license is required for multi-tenant SaaS-style redistribution.
• Backend API for integration, and MCP support via a published MCP Server URL (which carries authentication credentials).
• Free self-hosted; Dify Cloud is a paid SaaS with free and paid tiers.

What this means. Dify is the neutral counterweight in a vendor-bound ecosystem: a platform with no model of its own, whose entire value is orchestrating other vendors’ models behind a consistent workflow and RAG layer.

What is uncertain. The exact boundary of “multi-tenant SaaS redistribution” that triggers a commercial license, the company’s funding, production sizing beyond the 2-core/4 GiB minimum, and the independent output-quality of Dify-built applications are not recorded in this database.

Sources.

evidence_id	source_name	source_url	source_type	published	verified	confidence	conflict
src-agents-dify-1	Dify GitHub repository	https://github.com/langgenius/dify	Official documentation	—	2026-09-30	high	—
src-agents-dify-2	Dify official website	https://dify.ai/	Official	—	2026-09-30	high	—
src-agents-dify-3	Dify documentation	https://docs.dify.ai/	Official documentation	—	2026-09-30	high	—
src-agents-dify-4	Dify LICENSE	https://github.com/langgenius/dify/blob/main/LICENSE	Official documentation	—	2026-09-30	high	—
src-agents-dify-5	Dify MCP Server documentation	https://docs.dify.ai/en/cloud/use-dify/publish/publish-mcp	Official documentation	—	2026-09-30	high	—
src-agents-dify-6	Dify pricing	https://dify.ai/pricing	Official	—	2026-09-30	high	—

Why it matters

Dify matters as the reference implementation of model-neutrality as a product. Where Coze defaults to Doubao, Baidu AppBuilder to ERNIE and Yuanqi to Hunyuan, Dify has no default at all — it orchestrates hundreds of models behind one workflow and RAG layer.

China AI Hub analysis indicates Dify’s structural role is orchestration-without-a-model: it sells the workflow, retrieval and observability layers, not any model, which is both its strength (no lock-in) and its cost (the operator must assemble and pay for model access themselves). That positions it as the escape hatch from the bound platforms, and the research on the agent ecosystem structure records exactly this contrast — most Chinese agents are bound to their vendor’s models; Dify is the opposite.

China AI Hub analysis: Dify’s openness is more consequential than a bound lab agent’s, but it narrows in one specific place — the license. It stays free to run but not free to resell as a competing multi-tenant SaaS, a boundary that protects the Dify Cloud business model. That is the open-core pattern in miniature: the self-hosted core is free, and the commercial boundary is drawn exactly where a reseller could undercut the managed service.

How it differs from FastGPT and MetaGPT

Dify, FastGPT and MetaGPT are all open-source and model-agnostic, but they sit in different cells of the ecosystem.

• FastGPT is knowledge-base-first: document ingestion, chunk management, hybrid retrieval and rerank are its primary surface, aimed at enterprise Q&A over internal corpora.
• MetaGPT is a framework, not a platform — a pip-installable, role-based multi-agent SDK that a developer composes into their own system, with a commercial product (MGX) layered on top.
• Dify is a platform — a self-hostable application builder spanning workflow, RAG, agents and observability, oriented to prototype-to-production application shipping.

China AI Hub analysis: the three are complementary rather than competing in the same lane. Dify is the general-purpose application platform; FastGPT is the specialist for retrieval-heavy knowledge work; MetaGPT is the idea-and-pattern source whose commercial energy has shifted to a productized layer. A team choosing Dify is choosing breadth of application scope over depth in any single area — the reverse of FastGPT’s bet.

Practical implications

For teams escaping lock-in. Dify is the documented, self-hostable path to run one workflow across many models — the strongest fit in the database for a team that wants to swap models without rebuilding its application.

For self-hosters. The trade-off is real infrastructure: a 2-core CPU / 4 GiB RAM minimum, and the operator must configure and pay for every model provider. Dify gives data and routing control in exchange for operational responsibility — the inverse of a bound platform’s managed convenience.

For MCP publishers. Dify exposes applications as MCP Servers, but the published URL carries authentication credentials — a security consideration that bound platforms (whose MCP is internal) do not surface the same way.

For those who might resell. The non-OSI license is the key legal gate: self-hosting for internal use is free, but redistributing Dify as a multi-tenant SaaS requires a commercial license from LangGenius.

What the evidence shows

The evidence for Dify is strong because it is a mature open-source project with a public repo, docs, LICENSE and a pricing page. The README’s own framing — “AI workflow, RAG pipeline, agent capabilities, model management, observability features” — grounds the “orchestration-without-a-model” reading in the vendor’s primary material, and the LICENSE text confirms the commercial-redistribution condition that the non-pure Apache-2.0 label flags.

The gaps are licensing precision and independent evaluation. The exact trigger for the commercial license is defined by the Dify Open Source License, not a standard OSI license, so the boundary is vendor-drawn; the company’s funding is not public; and no independent benchmark of Dify-built application quality is recorded here. China AI Hub analysis indicates Dify’s transparency is otherwise the highest among the agent platforms in this database, precisely because it is open — the repo and license are inspectable in a way the closed platforms’ pricing and model routing are not.

Where this fits

Workload	Relevance
Agentic workflow + RAG application building	High
Prototype-to-production LLM application shipping	High
Cross-model orchestration (hundreds of providers)	High
Observability and evaluation of LLM apps	High
Self-hosted / VPC deployment for data control	High
Knowledge-base-first enterprise Q&A	Moderate (FastGPT is deeper)
Reselling as a multi-tenant SaaS	Restricted (commercial license)

Relevance judgments are China AI Hub analysis based on documented capabilities, not vendor claims.

Field reference

Field	Value	Evidence type
Underlying model	Model-agnostic (hundreds of LLMs; any OpenAI API-compatible model)	Official
Target users	Developers, application teams, self-hosters	Official
Platform	Web app + Docker self-hosted; cloud SaaS	Official
OS	Cross-platform (Docker); Linux server	Official
Browser / computer use	Not publicly documented	Not publicly documented
Coding	Not a primary focus (application workflow, not a code agent)	Official
Autonomous task execution	Yes (agent capabilities, workflow)	Vendor-reported
MCP	Yes (MCP Server publish)	Official
Tool calling	Yes (workflow tools, agent tools)	Official
Memory	Yes (RAG, conversation memory)	Official
Workflow	Visual workflow canvas + RAG pipeline + agent layer	Official
API	Yes (backend service / API)	Official
Pricing	Open source (free); Dify Cloud paid SaaS tiers	Vendor-reported
Region	Global (Singapore company; cloud + self-hosted)	Official
Open-source	Yes — Apache-2.0 based (Dify Open Source License)	Official
Deployment	Both (cloud + self-hosted)	Official
Limitations	Non-pure Apache-2.0 (SaaS needs commercial license); 2-core/4GiB minimum	Official
Source	GitHub · dify.ai · docs	Official
Last verified	2026-09-30	Official

See the LangGenius company profile, the self-hosting guide, the open-weight vs API economics, and the site’s AI agents, RAG and MCP technology pages. For the structural reading of Dify as the neutral counterweight to the bound platforms, see the research on the agent ecosystem structure and the rise of Chinese AI agents.

Labels used above: Official fact (from the Dify GitHub repository, LICENSE, documentation and pricing page), Vendor-reported claim (feature and deployment statements by LangGenius), and China AI Hub analysis (our synthesis, always introduced as such). No third-party evaluation evidence is currently recorded for Dify-built applications.