hekmon
c4f2fb1dfd
The TGI repository was archived on 2026-03-21 and is in maintenance mode since late 2025; Hugging Face now recommends vLLM and SGLang as forward inference engines.
ComfyUI OpenAI API
This repository contains ComfyUI nodes that integrates with the OpenAI API: it allows you to use language models and vision language models within your workflow.
It is KISS by design and intended for those who only wants basic capabilities without having to import massive projects like LLM party: only the chat completions endpoint (with vision support) is implemented as it should be enough for 99% of use cases.
Thanks to its simplicity the project has a low footprint: it only has 1 external dependency (3 in total) !
openaithe official OpenAI API bindings for Pythonnumpyfor computation but it is already a dependency of ComfyUIPillowalso already used by ComfyUI for image processing
Usage
The default base_url parameter value targets the official OpenAI API endpoint by default but by changing it, you can also use this project with any OpenAI API compatible servers like Ollama, vLLM, SGLang, etc...
Multiples images are supported as long as they are fed batched to the chat completion node.
If you want to customize the chat completion, you can chain options to modify the request. Most common options are available as predefined nodes but you can inject any key/value pair using the Extra body node.
Options nodes are available for:
seedtemperaturemax_tokensmax_completion_tokenstop_pfrequency_penaltypresence_penaltydeveloper_roleextra_body(for any other key/value pair)
Use either max_tokens or max_completion_tokens, not both. Older OpenAI-compatible servers keep using max_tokens. Newer OpenAI chat models that reject max_tokens use max_completion_tokens.
API key
The client node reads the API key from the OPENAI_API_KEY environment variable at execution time. It is intentionally impossible to enter an API key inside the workflow.
Never store a real API key in a workflow. ComfyUI serializes every node widget value — even from bypassed nodes — into the prompt and workflow metadata chunks of generated images and exported workflow files. Anyone you share an image or a workflow with (Discord, Reddit, forum attachments...) can read the key back from those metadata. This is a ComfyUI core behavior (see the official workflow metadata documentation) which no custom node can opt out of: any node exposing a key widget leaks it.
How to configure:
- Official OpenAI API or any authenticated endpoint: set the environment variable before starting ComfyUI, e.g.
export OPENAI_API_KEY=sk-...(Linux/macOS) orsetx OPENAI_API_KEY sk-...then restart your shell (Windows). In Docker, pass it with-e OPENAI_API_KEY=sk-.... - Keyless inference servers (vLLM, some Ollama setups, etc.): those servers require an
Authorizationheader but do not validate its value. If the environment variable is not set, the client automatically falls back to the-placeholder, so nothing to configure. - Defense in depth: if you generate images to share, launch ComfyUI with
--disable-metadata(or enable Settings > Server-Config > Disable saving prompt metadata in files in ComfyUI Desktop) to strip workflow metadata from output files entirely.
If you have used a version of this project prior to v3.0.0 with a real key and shared generated images or workflow files, rotate that key and manually clean up your saved workflows: the previous versions stored the key inside every image and workflow export.
Installation
ComfyUI Manager
Search for OpenAI API in the Custom Nodes Manager and install it.
Manually
On the github interface, click the green <> Code button and then Download ZIP. Extract the root folder of the zip file into your ComfyUI/custom_nodes directory.
Example
