EmAySee Advanced Oobabooga Connector

ComfyUI runs diffusion models locally, and it can run a text model too - you just need a bridge. EmAySee Advanced Oobabooga Connector is that bridge: it sends your prompt to a text-generation-webui (Oobabooga) instance running an OpenAI-compatible API and returns the model's answer as a string you can feed anywhere in the graph.

The promise is real: one node, a prompt in, an answer out, and suddenly you have an LLM inside ComfyUI doing prompt rewriting, tag generation, workflow scripting, whatever you can get a local model to do. And "local" is the operative word - no API key from a cloud vendor, no per-token billing, the same spirit as running your own Stable Diffusion. The author, who posts about these nodes on Reddit, is upfront that this family of connectors is not a one-and-done: "you need to craft a good prompt to have it convert though," they wrote, "you need to give the LLM a set of instructions." Treat it as a raw API client, not a magic prompt-improver.

How it works

The node posts a JSON payload to your api_url using the OpenAI /v1/chat/completions schema, with system_prompt and prompt as the message pair. Around fifty sampler settings - temperature, top_p, top_k, min_p, repetition penalty, DRY, mirostat, and more - are forwarded if you've changed them from defaults, so it behaves like the full text-generation-webui API exposed as a node.

Two behaviors stand out. First, auto_continue: if the model stops because it hit max_tokens, the node re-calls the API up to max_continues times, appending the output until the model finishes naturally - a genuine convenience for long generations. Second, thinking parsing: it scans the response for <think>...</think> blocks (the format reasoning models like DeepSeek R1 use) and splits the output into answer and thinking. If a <think> tag is opened but never closed, the answer becomes the literal string [REASONING TRUNCATED] and the thinking gets the rest - a design choice worth knowing about, because that exact string is a sentinel some of the pack's other nodes watch for.

The inputs that matter

There are a lot of them, but a beginner sets these:

• api_url - defaults to http://10.0.0.71:5000/v1/chat/completions. That's the author's LAN address; you must change it to your own Oobabooga server. It must be running with its API enabled (--api flag) at the matching port.
• api_key - defaults to supersecretkey. If your server isn't enforcing auth, any value works; if it is, use the real one.
• prompt and system_prompt - the actual conversation. This is where you invest your effort.
• seed (-1 = random), max_tokens, temperature, top_p, min_p - the core generation controls.

Outputs are answer (clean text), thinking (the extracted reasoning), and full_raw (the entire JSON response, pretty-printed - your debugging friend).

Install

Part of the EmAySee pack:

cd ComfyUI/custom_nodes
git clone https://github.com/EmAySee/ComfyUI_EmAySee_CustomNodes

Restart ComfyUI (or Manager → "ComfyUI_EmAySee_CustomNodes"). No requirements.txt - it needs only requests, which ComfyUI already ships. You do need a running Oobabooga text-generation-webui with the API enabled.

Gotchas

The default api_url pointing at 10.0.0.71 is the #1 gotcha - every new user hits a connection error until they change it. Next, stop_on_error defaults to True, meaning a backend error crashes the workflow; flip it to False if you'd rather get an error string and keep going. Timeout is 600 seconds, so a slow model won't bail early - but it also means a hung server stalls your workflow for ten minutes. And remember: this is hobbyware with an explicit no-support disclaimer, and the "OpenAI-compatible" label means it should work with other compatible servers (vLLM, LM Studio) but is only tested against Oobabooga. Start with reasoning_effort: medium and a well-crafted system prompt, and you're most of the way there.