ComfyUI-API-Pack

ComfyUI-API-Pack

English | 한국어

Nodes that call external APIs: zero-shot classification on llama-server, OpenAI-compatible inference, remote ComfyUI API execution and Grok image-to-video.

Example

The zero-shot workflow asks three questions about a customer message. Zero-Shot Classification answers all of them in one request and prints one line per question, and Zero-Shot Answer picks out the route answer.

The inference workflow writes prompts with OpenAI Inference and generates an image from them.

The API workflow loads a workflow with Load Workflow and runs it on a remote ComfyUI, once with Api Generate and once with Api Submit and Api Collect.

Usage

Zero-Shot Classification

Write the labels in questions, feed the text into state, read the answers out of summary and answers. It calls /v1/systemone on a llama-server running a decision model such as clef-flash, which scores every label in one forward pass instead of generating text.

| Widget | What it does | |--------|--------------| | state | The content to classify | | questions | A JSON object mapping a question id to a typed question, see below | | base_url | llama-server base URL. Default http://localhost:8082 | | api_key | Sent as a Bearer token when set | | model | Sent when set. A server with one model ignores it | | image (input) | Images placed before the state, one per batch item. The server must be started with --mmproj | | answers (output) | The response's answers as indented JSON, keyed by question id | | summary (output) | One id: value (score) line per question, with value and score as in Zero-Shot Answer |

| Type | criteria | Answer | |------|------------|--------| | choice | {"option": "description" or null, ...} | choice, confidence, probabilities | | score | ["lowest", ..., "highest"], 2 to 10 levels | expected score, confidence, legend, probabilities | | noul | Optional {"true": "description", "false": "description"} | noul, the probability of true |

{
  "route": {
    "type": "choice",
    "instructions": "Which team should handle the message?",
    "criteria": {"billing": "Payments or invoices", "technical": "Bugs or outages"}
  },
  "urgency": {"type": "score", "instructions": "How urgent is it?", "criteria": ["Can wait", "This week", "Today"]},
  "outage": {"type": "noul", "instructions": "Is a service down?"}
}

With the state Our checkout started returning errors and orders are blocked., summary is:

route: technical (0.96)
urgency: Today (1.88)
outage: true (0.81)

The same input always gives the same answers. There is no temperature, top_k, top_p or seed: nothing is sampled, and the model file fixes how the probabilities are scaled.

OpenAI Inference

Feed a prompt into OpenAI Inference, read the answer out of response. Works with any OpenAI-compatible backend: OpenAI, vLLM, Ollama's /v1 endpoint, Gemini's OpenAI-compatible endpoint.

| Widget | What it does | |--------|--------------| | prompt / system_instruction | User and system prompts | | base_url | API base URL, e.g. https://api.openai.com/v1. Empty reads OPENAI_BASE_URL | | api_key | Empty reads OPENAI_API_KEY | | model | Empty picks the model from /models when the server has exactly one | | max_output_tokens | Up to 131072. Default 100 | | seed / temperature | Sampling. temperature 0.0 to 2.0, default 0.7 | | think | Turns on thinking mode | | image (input) | Input image for vision models | | response (output) | The answer, with any <think> block stripped out | | reasoning (output) | The thinking trace from reasoning_content or an inline <think>...</think> block, empty if none |

The last 10 distinct requests are cached in memory, so running an identical request again does not call the API.

Api Generate / Api Submit / Api Collect

Run a workflow on a remote ComfyUI over its HTTP API, such as a RunPod pod. The workflow is the API-format JSON (ComfyUI menu: "Save (API Format)"), not the UI format. Api Generate waits for the result; Api Submit returns a job_id at once and Api Collect picks the result up later.

| Widget | What it does | |--------|--------------| | workflow (input) | API-format workflow JSON, or a path to one. Load Workflow reads one from user/default/workflows/ | | api_url | Remote ComfyUI base URL, e.g. https://xxxx-8188.proxy.runpod.net/ or http://127.0.0.1:8188 | | positive_prompt (input) / positive_prompt_id | The prompt and the node id that receives it | | negative_prompt (input) / negative_prompt_id | Same for the negative prompt. Skipped when empty | | seed / seed_id | The seed and the node whose seed/noise_seed receives it. -1 keeps the workflow's seed | | image (input) / image_node_id | An image uploaded to the remote and bound to that LoadImage node | | output_id | The node whose images/gifs output is fetched. Animated outputs become frames | | overrides (input) | JSON {node_id: <full node dict>}. Each entry replaces the whole node, applied last | | timeout_sec | Api Generate only. Polling limit in seconds, default 300 | | label | Api Submit only. Recorded with the job and returned by Api Collect | | wait_sec / poll_interval | Api Collect only. wait_sec 0 collects if ready and otherwise skips at once |

Api Collect returns the frames when the job is done. Otherwise it returns an ExecutionBlocker, which skips the nodes downstream.

Grok Generate / Grok Submit / Grok Collect

Turn one image into a Grok Imagine video clip, saved to the output folder as a native VIDEO with an inline preview. Grok Generate waits for the clip; Grok Submit returns a request_id at once and Grok Collect picks the clip up later.

| Widget | What it does | |--------|--------------| | image (input) | The first frame | | prompt | Motion or scene description, optional | | duration / resolution | 1 to 15 seconds, 720p or 480p | | model | Default grok-imagine-video-1.5-preview | | filename_prefix | Output path under the ComfyUI output directory. Default grok/GrokVideo | | poll_interval / timeout | Grok Generate polling. Defaults 5 and 600 seconds | | access_token / refresh_token / client_id | Empty reads GROK_ACCESS_TOKEN / GROK_REFRESH_TOKEN / GROK_CLIENT_ID. The access token refreshes itself on a 401/403 | | label | Grok Submit only. Recorded with the job and returned by Grok Collect | | wait_sec / poll_interval | Grok Collect only. wait_sec 0 collects if ready and otherwise skips at once |

Grok Collect calls the Grok API too, so it needs the tokens as well. They are not stored with the job.

The API and Grok nodes keep one in-flight job each in jobs.lock in the pack directory. Submit skips while a job of its kind is in flight, and Collect frees the slot once it finishes. Run Collect in a loop (e.g. /loop) to pick up the result when it is done.

Installation

Search for ComfyUI-API-Pack in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/alchemine/comfyui-api-pack
pip install -r comfyui-api-pack/requirements.txt

Nodes

ApiPack/Inference

Zero-Shot Answer: picks one question out of the Zero-Shot Classification answers.

| Output | choice | score | noul | |--------|----------|---------|--------| | value (STRING) | The chosen option | The description of the nearest level | "true" if the probability is 0.5 or more, else "false" | | score (FLOAT) | Probability of the chosen option | Expected level | Probability of true | | confidence (FLOAT) | confidence | confidence | \|2p - 1\|, as for a two-option choice |

ApiPack/API

Load Workflow: reads an API-format workflow JSON from user/default/workflows/ as a STRING.

Configuration

.env is optional. Copy .env.example to .env and set only what you use; the node inputs take priority over it. A missing credential raises an error when the node runs, not when the pack loads.

OPENAI_BASE_URL=https://api.openai.com/v1
OPENAI_API_KEY=your-api-key

GROK_ACCESS_TOKEN=...
GROK_REFRESH_TOKEN=...
GROK_CLIENT_ID=...

A Grok token refresh failed (...) error means the refresh token or client_id has expired or been revoked: re-authenticate with x.ai and update them.

License

GPL-3.0. See LICENSE.