Migrate from Dify to Heym
Move Dify apps to Heym. Paste an exported workflow, watch AI Convert rebuild every node on a live canvas, then run it on infrastructure you control.
Dify presents itself as an LLM application development platform where teams build chat apps, workflows, and agents from a node-based editor.
When it is worth moving a workflow
The usual trigger is an app outgrowing its own boundary. Once an assistant has to reach into internal systems on a schedule, open a pull request, drive a browser, or stop for an approval before it acts, the work is no longer only an LLM application. Moving it onto a general workflow runtime keeps the model logic and puts the operational half of the job on the same canvas.
If you are still deciding rather than moving, the Heym and Dify comparison covers the evaluation side with links to first-party documentation.
How the move works
- 1
Export from Dify
Open the app in Dify, export it as a DSL file from the app menu, and copy the file contents.
- 2
Paste it into AI Convert
A safety check runs, then each construct is mapped to the closest native Heym node and streamed onto a live canvas while you watch.
- 3
Review, download, and import
Read the rebuilt workflow, download or copy it, import it into your own instance, and attach your credentials.
What each Dify concept becomes
A reviewer's guide to the rebuilt workflow. Constructs without an exact equivalent are mapped to the closest Heym primitive, which is why the result is worth reading before it handles real traffic.
| In Dify | In Heym | What changes |
|---|---|---|
| Start node | Input node | Declared input variables become the workflow input schema. |
| LLM node | LLM node | Prompts, model selection, and structured JSON output carry over, with guardrails available on the same node. |
| Knowledge Retrieval | RAG / Vector Store node | A dataset becomes a vector store in the Vectorstores tab, backed by Qdrant or Postgres with pgvector. |
| Question Classifier | LLM node with JSON output, then a Switch node | Classification and routing become two explicit steps, so the label the model chose is visible in the run. |
| IF/ELSE | Condition node | Branch conditions carry over to the same true and false paths. |
| Iteration | Loop node | Per-item processing is preserved, and independent items can run in parallel. |
| Variable Aggregator | Merge node or Variable node | Collecting several branch outputs becomes a Merge; carrying a value across a run becomes a Variable. |
| Code and Template nodes | Code node and Set node | Python transformations move into the sandboxed Code node; string templating becomes an expression in a Set field. |
| HTTP Request | HTTP node | Requests carry over with the same methods, headers, and body templating. |
| Agent node and tools | AI Agent node | Tool use carries over, extended with inline Python tools, MCP servers, skills, and sub-agent delegation. |
| Answer and End nodes | Output node | The final response becomes the workflow output, available to the API, the Portal, and any calling workflow. |
| Published web app | Portal | A workflow publishes as a chat UI with streaming, file upload, conversation history, and optional authentication. |
Check these before it goes live
- Rebuild the datasets. Documents are re-ingested into a Heym vector store rather than copied, so upload them and confirm retrieval quality against a few known questions.
- Attach model credentials in the Credentials tab and confirm the model each LLM node points at.
- Check the conversation surface. If the app was published as a Dify web app, publish the Heym workflow through the Portal and re-test multi-turn behavior.
- Verify the classifier path. Classification and routing are two nodes now, so confirm each label reaches the branch you expect.
- Run one test execution and compare its answer against the same input on the source side.
Dify migration FAQ
Can AI Convert read a Dify export?
Yes. Paste the exported app DSL into AI Convert and it maps each node to the closest native Heym node, streaming the rebuilt workflow onto a live canvas you can download and import.
Do my knowledge base documents move with the app?
The workflow structure moves; the documents are re-ingested. Create the vector store in the Vectorstores tab, upload the same sources, and check retrieval against a handful of questions you already know the answers to before switching traffic over.
Do my credentials come across?
No, and that is deliberate. Exported workflow files hold no API keys or secrets, so AI Convert rebuilds the structure only and you attach your own credentials in the Heym Credentials tab afterwards. Nothing copies a secret between two systems on your behalf.
Is the converted workflow ready to run?
Treat it as a reviewed starting point rather than a finished port. Open it on the canvas, attach credentials, run one test execution, and adjust any node that was mapped to the closest Heym primitive rather than an exact equivalent.
Can I run the result on my own infrastructure?
Yes. Heym is source-available under an MIT license with a Commons Clause and runs on your own servers with Docker or Kubernetes, so migrated workflows, their credentials, and their model calls stay inside your environment.
Moving from something else
AI Convert reads exports from 12 tools, and it also accepts a plain description of the automation when there is no file to export.