Description
n8n-nodes-stac
This is an n8n community node package. It lets you use Stac as a language model provider in your n8n workflows, and extract data from documents or generate PDFs with it.
Stac gives you access to open-source LLMs for a fixed monthly price, through an OpenAI-compatible API. Instead of paying per token, you pay a flat subscription — under saturation Stac degrades speed rather than cutting you off.
n8n is a fair-code licensed workflow automation platform.
Installation
Operations
Credentials
Compatibility
Usage
Resources
Version history
Installation
Follow the installation guide in the n8n community nodes documentation, using the package name n8n-nodes-stac.
Operations
This package provides two nodes.
Stac Chat Model
Stac Chat Model — a language model that plugs into the Model input of any n8n AI root node (AI Agent, Basic LLM Chain, Question and Answer Chain, and so on).
Parameters:
| Parameter | Description |
| — | — |
| Model | Loaded dynamically from your Stac plan via GET /models, so the dropdown only lists models you can actually call |
| Sampling Temperature | 0–2. Lower values make responses more deterministic |
| Maximum Number of Tokens | Upper bound on tokens generated in the response. When the option isn’t added, the node sends 2048 — or nothing (the server’s floor of 8000 applies) when Reasoning Effort is low, medium or high. With reasoning on, values below 8000 can cut the reasoning off mid-way |
| Reasoning Effort | none, low, medium or high, sent as reasoning_effort. When the option isn’t added, the reasoning level configured on your API key applies |
| Timeout | Request timeout in milliseconds (default 120000). Under saturation Stac degrades speed, so aggressive values can cut off otherwise valid responses |
| Max Retries | How many times to retry a failed request (default 2) |
Stac
A regular node (input and output on the canvas) that covers the Stac API endpoints, grouped like n8n’s OpenAI node:
| Resource | Action | Endpoint |
| — | — | — |
| Text | Message a model | POST /chat/completions |
| Image | Analyze image | POST /chat/completions with image input |
| | Edit image | POST /images/edits (image plan) |
| | Extract data from an image | POST /images/extract |
| | Generate an image | POST /images/generations (image plan) |
| Document | Extract data from a document | POST /documents/extract (or /images/extract for an image) |
| | Generate a PDF | POST /documents/generate |
| Decision | Make a decision | POST /decisions |
| Model | List models | GET /models |
Message a model — a list of messages (user, assistant, system; a system message replaces the stack’s system prompt). Model is loaded from GET /models; leave it on Stack Default Model to use your stack’s own. Options: Maximum Number of Tokens (not sent unless added — the server’s 8000 floor applies), Output JSON (responseformat: jsonobject), Reasoning Effort (none/low/medium/high; not added = the key’s level), Sampling Temperature, Top P. Simplify Output returns content and finish_reason; with Output JSON on, content is parsed into an object.
Analyze image — a question plus one or more images, from binary fields (comma-separated) or URLs. URLs are downloaded by n8n and sent inline as data URIs, so the image only has to be reachable from n8n. Same model and options as Message a model.
Extract data (document or image) — PDF or image in, JSON validated against your JSON Schema out. OCR runs automatically (scanned PDF pages; every image). Put every field in required and use nullable types (["string", "null"]) on the ones that can be missing, or the model invents a value. Several files in one extraction: comma-separated binary fields (Pro plan and up). Simplify Output returns only the extracted fields. Options: Context, System Prompt, Max Tokens (default 4000).
Generate an image / Edit image — needs a key from a stack on the image plan. Pick the image model in Model (loaded from GET /models). The PNG goes into a binary field and the JSON carries meta (effective size, steps, seed…) and usage. Edit takes up to 4 reference images (comma-separated binary fields). Options: Size (1024x1024, 1536x1024, 1024x1536), Steps (1–8), Guidance Scale (0–20), Seed (kept as text, so the full 0–2^64−1 range survives), File Name. An empty prompt uses the prompt configured on the key; prompts are cut at 512 tokens.
Generate a PDF — from ready-made HTML (no model involved, no tokens spent) or from an instruction the model turns into HTML. HTML must be self-contained: external URLs are never fetched — use data: URIs.
Make a decision — typed answers from a decision model (Jev, by TypeSafe) instead of free text. State is a JSON object describing the situation (build it from the item with an expression); Questions is a JSON object of named questions, each with a type (choice, noul or score), instructions and criteria. The response carries one answer per question with its choice, confidence and probabilities. Model is a text field (default typesafe/jev-1.13) because GET /models doesn’t list decision models; a model outside your plan comes back as 403, and an unknown one as 404 listing the accepted ones. Simplify Output returns only the answers; off, it also returns usage.
All operations wait up to 5 minutes for the API. API errors surface with the API’s own message and HTTP status (for example, 403 when an image action runs with a text-plan key).
Credentials
You need a Stac account and an API key.
1. Sign up at trystac.com and copy your API key.
2. In n8n, create a new Stac API credential.
3. Paste the API Key. Leave Base URL at https://api.trystac.com/v1 unless you were given a different endpoint.
Authentication uses a bearer token (Authorization: Bearer ). The credential’s test button calls GET /models, so a successful test confirms both the key and the base URL.
Compatibility
Developed against n8n 2.43.7 (@n8n/ai-node-sdk 0.33.5); versions up to 1.1.0 were tested against n8n 2.36.9.
The node builds the model through supplyModel from @n8n/ai-node-sdk, which n8n provides at runtime — the package declares @n8n/ai-node-sdk and n8n-workflow as peer dependencies and bundles no LangChain code of its own. This keeps a single @langchain/core instance in the process and means model executions show up in n8n’s own log/tracing view. Older n8n releases ship earlier versions of that SDK and have not been tested.
Usage
Stac Chat Model: add an AI Agent (or any AI root node) to your canvas, then connect Stac Chat Model to its Model input. Pick a model from the dropdown — if the list is empty, your credential is wrong or the base URL is unreachable.
Stac: pick an action from the node panel. For the file-based actions (extract, analyze from binary, edit image), put the node after one that outputs a file — Read/Write Files from Disk, HTTP Request with a file response, a form or email trigger with attachments. Actions that produce a file (generate image, edit image, generate PDF) put it in a binary field for the next node to store or send.
New to n8n? Start with the Try it out documentation.
Resources
Version history
1.5.0
New Decision → Make a decision action on the Stac node, for the Jev typed-decision model (POST /decisions).
1.4.1
Generate an image and Edit image get a required Model field, loaded from the image key’s GET /models and sent as model.
1.4.0
The Stac node now covers every Stac endpoint, grouped like the OpenAI node: Text (message a model), Image (analyze, edit, extract data, generate), Document (extract data, generate PDF) and Model (list). Workflows built on 1.3.0 keep working.
1.3.0
New Stac node with Extract Data (PDF/image → JSON, with OCR) and Generate PDF. Icon files renamed so n8n stops serving the old logo cached under the previous file name.
1.2.0
Adds the Reasoning Effort option and stops sending the 2048 token fallback when reasoning is requested, so the reasoning step isn’t truncated. Dev dependencies aligned with n8n 2.43.7.
1.1.0
New Stac logo (light and dark variants), Maximum Number of Tokens defaults to 2048 even when the option isn’t added, and streaming responses with usage. Tagged but never published to npm — its changes ship in 1.2.0.
0.1.0
Initial release: Stac Chat Model sub-node with dynamic model loading, plus the Stac API credential.
