Original video in Chinese.
Key Takeaway
- Dify has added a “bidirectional MCP” feature, allowing users to add MCP servers inside Dify and also convert Dify workflows into MCP servers for external output.
- This feature solves the workflow “island” problem, enabling workflows to integrate into everyday general-purpose scenarios and greatly expanding Dify’s tool range.
- I’ve started paying attention to Dify again, because workflows should be atomized and used as components to strengthen results in specific scenarios.
- Through a demo of the Deep Research workflow, this article shows how, after combining Dify with MCP, AI clients can call specific MCP servers to solve specific needs and improve efficiency.
- Combined with prompts, Dify’s MCP feature can achieve workflow automation based on MCP tools and prompts, improving personal productivity.
Dify recently launched a new feature: bidirectional MCP. This is a very important feature. Because it allows workflows that were originally aimed only at specific scenarios to merge into our daily, general-purpose usage scenarios.
What does “bidirectional MCP” mean?
One direction is inbound, meaning you can add existing MCP servers inside Dify. The benefit is that it can greatly expand Dify’s tool range and incorporate an increasingly rich set of MCP servers.
The other direction is outbound, meaning you can convert the workflows you create into MCP servers and output them externally. For example, I use an AI client and add the workflow MCP into it. Then I can directly call it during conversations.
In this way, the workflows you spent so long building are no longer limited to certain usage scenarios and situations; they’re no longer islands. This is a huge improvement.
Hello everyone, welcome back to my channel. Modestly speaking, I’m one of the few creators in China who can clearly explain the Why and How of AI. What I provide is far more valuable than tutorials. Remember to hit follow. If you want to connect with me, come join our newtype community. The community has been running for 500 days, and more than 1,500 friends have joined as paid members.
Back to today’s topic: Dify’s MCP feature.
I remember making a video introducing Dify last year. But after that, I stopped paying attention to these workflow platforms. Two reasons:
First, the barrier to entry is high. These workflows look like you can just casually assemble them on a canvas, but in fact they’re pretty difficult. You need to understand both technology and business. How many people in a company can reach that level?
Second, the use cases are narrow. Each workflow exists for certain scenarios and to solve specific problems. In the early days, this was very useful, because large language models weren’t very capable yet, so we organized human experience into workflows to guide the model. But today, if you still force a fixed workflow onto the model, it starts to constrain it.
So why am I picking Dify back up now, and why am I paying attention to this update?
Because workflows are not getting bigger; they’re getting smaller. They can become atomized components, integrated into my workflows to strengthen results in specific scenarios. That is the value an MCP server should have.
Let me show you, and you’ll understand.
This is a ready-made workflow I found in Dify: Deep Research. I changed the model inside to GPT-4.1, then published it. Because only after publishing can you enable the MCP server feature.
Click this button on the left, and in the lower-left corner turn on this option, and you can convert this workflow into an MCP server. This line is the server address. Since I’m running it locally, the address starts with localhost.
Open the AI client. I’m using the free Cherry Studio here. Create a new MCP. Choose HTTP for the connection method. Then fill in the server address, and you’re done.
Let’s make a comparison.
First, I use the model’s built-in search tool to search for a question: What is Context Engineering? This is its answer.
Then I open a new window. This time I use the Deep Research MCP server I just connected. Same question: What is context engineering?
After a few minutes, the model gave its answer. Comparing the outputs from the two runs, you can clearly see that using an external MCP server works much better.
You see, this is what I just said: when you reach a specific scenario and specific need, you call a specific MCP server to satisfy it and solve it. You don’t need to switch tools; you still use the original AI client, and that’s really convenient.
I used Deep Research for the demo because I wanted it to be easy to understand. In fact, workflows are highly customized. Next, I’ll build a number of workflows tailored to my own needs, and then convert them all into MCP servers.
So, as I said in the community, there are now two systems that can meet our customization needs: one is prompts, and the other is MCP servers based on workflows. And these two systems can also be unified and orchestrated by the product I made — Prompt House.
With this setup, I believe my personal productivity will see another big improvement.
OK, that’s it for this episode. If you want to learn about AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you next time!