Migrate from Langflow to Heym
Move Langflow flows to Heym. Paste an exported flow JSON, watch AI Convert rebuild each component on a live canvas, then run it on your own infrastructure.
Langflow describes itself as a visual builder for LLM applications, where flows are assembled from components for models, prompts, retrieval, tools, and agents.
When it is worth moving a workflow
A flow that answers well still has to become something operable. The move usually happens when the same graph needs a schedule, an approval before it acts, a record of what each run cost, and integrations that reach past the model into the systems the answer is about.
If you are still deciding rather than moving, the Heym and Langflow comparison covers the evaluation side with links to first-party documentation.
How the move works
- 1
Export from Langflow
Open the flow in Langflow and export it as JSON from the flow menu, then 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 Langflow 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 Langflow | In Heym | What changes |
|---|---|---|
| Flow | Workflow | Both tools think in nodes and edges, so the graph shape survives the move almost intact. |
| Chat Input and Chat Output | Input node and Output node | The conversational surface itself moves to the Portal, which publishes the workflow as a chat UI. |
| Prompt component | Prompt field on the LLM or Agent node | A prompt stops being a separate component and becomes a field with expressions on the node that uses it. |
| Model component | LLM node | Model, temperature, and structured JSON output carry over, with guardrails available on the same node. |
| Agent component | AI Agent node | The tool-calling loop carries over, extended with inline Python tools, MCP servers, skills, and sub-agent delegation. |
| Tool components | Python tools and MCP servers on the Agent node | Tools are configured on the agent itself rather than wired in as separate nodes. |
| Embeddings and Vector Store components | RAG / Vector Store node and the Vectorstores tab | Ingestion collapses onto one node over Qdrant or Postgres with pgvector; the documents are re-ingested rather than copied. |
| Retriever component | RAG / Vector Store node in search mode | Semantic search with metadata filters and reranking, feeding straight into an LLM or Agent node. |
| Conditional Router | Condition or Switch node | Two-way checks become a Condition; multi-way routing becomes a Switch with labelled branches. |
| Loop component | Loop node | Per-item processing carries over, and independent items can run concurrently. |
| Custom Python component | Code node | Already Python, so this is the cleanest part of the move; it runs in a Docker sandbox here. |
| API endpoint and Playground | Workflow API and the Portal | Every workflow is callable over the API, and an agent workflow is also reachable as an MCP server. |
Check these before it goes live
- Re-ingest the documents. Vectors are rebuilt rather than copied, so load the same sources and spot-check retrieval.
- Attach model and tool credentials in the Credentials tab.
- Re-test multi-turn behavior through the Portal, since conversation history is held by the chat surface.
- Confirm each prompt landed on the node that uses it, now that prompts are fields rather than components.
- Run one test execution and compare the answer against the same input in the source flow.
Langflow migration FAQ
Can AI Convert read a Langflow export?
Yes. Paste the exported flow JSON into AI Convert. Langflow graphs translate unusually cleanly because both tools model work as nodes and edges rather than as a linear step list.
Do my custom Python components carry over?
The logic does, into the sandboxed Code node, which is also Python. Imports and any Langflow-specific base classes need adjusting, so read the generated node before it handles real traffic.
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.