src/bundles/openai-compatible/README.md
Generic OpenAI-compatible endpoints as a first-class Langflow model provider, packaged as a standalone Extension Bundle.
Registers an OpenAI Compatible model provider that appears in Langflow's unified model picker alongside the built-in providers. Any service that speaks the OpenAI HTTP API shape can be configured through it — for example:
| Provider | Base URL |
|---|---|
| OpenRouter | https://openrouter.ai/api/v1 |
| Together AI | https://api.together.xyz/v1 |
| Groq | https://api.groq.com/openai/v1 |
| Fireworks AI | https://api.fireworks.ai/inference/v1 |
| Self-hosted vLLM / TGI / LM Studio | http://localhost:8000/v1 |
The provider reuses the ChatOpenAI / OpenAIEmbeddings classes and discovers
the served models live from the endpoint's /v1/models route — the same model
is therefore offered in both the Language Model and Embedding Model
contexts.
Unlike a component bundle, this ships no component: it contributes a
provider through the providers[] block in extension.json, which Langflow's
provider registry merges into the unified model system. It edits no Langflow
core files.
Set the endpoint (and, if it requires one, an API key) under Settings → Model Providers → OpenAI Compatible, or via environment variables:
OPENAI_COMPATIBLE_BASE_URL — base URL of the endpoint, e.g.
https://openrouter.ai/api/v1 or http://localhost:8000/v1 (required).OPENAI_COMPATIBLE_API_KEY — bearer token, if the endpoint enforces one
(optional; local servers without auth do not need it).The provider holds one endpoint at a time. For a second concurrent custom
endpoint, combine it with the built-in OpenAI provider's Base URL override or
the vLLM provider (lfx-vllm).
pip install lfx-openai-compatible
The bundle is registered automatically via the langflow.extensions
entry-point. Restart your Langflow server and select OpenAI Compatible in
any Language Model or Embedding Model field.
cd src/bundles/openai-compatible
pip install -e .
lfx extension validate src/lfx_openai_compatible
Resolves the feature request in
#12839. Follows the
provider-bundle pattern established by lfx-vllm
(#13919).