Connect CNAPS.ai to n8n — MCP Integration Guide

CNAPS Studio exposes its entire platform as an MCP server. Point an n8n AI Agent at it and the agent can pick models, upload files, build flows, and run them — all from a single chat message, with no glue code.

This tutorial builds a working example end to end. When you finish, you'll be able to type this into an n8n chat window:

Upscale this image 4x: https://picsum.photos/id/237/320/240.jpg

and get back a finished, upscaled image.

Time required: about 10 minutes.

What you'll build

Four nodes:

Before you start

1. A CNAPS Studio API key. Create one at cnaps.ai under API Key.

2. An LLM API key for the agent's reasoning model — an Anthropic key for Claude, or an OpenAI key.

n8n's free trial AI credits only cover OpenAI (they appear as a separate n8n free OpenAI credits credential type). To use Claude, bring your own Anthropic key.

3. n8n 1.104.0 or later. The MCP Client Tool node needs a Server Transport parameter, which arrived in that release. Open the node and look: if you only see a field named SSE Endpoint and no transport selector, your n8n is too old to connect — upgrade first. The CNAPS.ai MCP server speaks streamable HTTP only and has no SSE endpoint, so there is nothing for an SSE-only node to connect to.

We tested this tutorial on n8n Cloud 2.34.4.

Step 1 — Create the workflow and add a chat trigger

Create a new workflow, then click Add first step.

Search for chat.

Search for chat, not chat message. Searching chat message returns only community nodes and you'll think the trigger is missing.

Pick Chat ("Runs the workflow when an n8n generated webchat is submitted"), then On new Chat event.

The node lands on the canvas labelled When chat message received, and a Chat panel opens at the bottom of the screen. That panel is how you'll test.

Step 2 — Add the AI Agent

Click the + on the right edge of the trigger, search AI Agent, and add the first result.

The agent node has three sub-connectors underneath it: Chat Model, Memory, and Tool. You'll use the first and the third.

Step 3 — Connect a chat model

Click the + under Chat Model.

Using Claude

Search Anthropic and add Anthropic Chat Model. The default model works well for this — leave it as is.

Click Set up credential and paste your Anthropic API key.

Set Allowed HTTP Request Domains to api.anthropic.com. It defaults to All, which lets any node using this credential send your key to any domain. Then save.

Using OpenAI

Click the + under Chat Model, choose OpenAI, and either use n8n's free trial credits or your own key.

Step 4 — Add the MCP Client Tool

Click the + under Tool, search MCP Client, and add MCP Client Tool.

Configure it exactly like this:

Endpoint:            https://mcp.cnaps.ai/mcp
Server Transport:    HTTP Streamable
Authentication:      Header Auth
Tools to Include:    Selected

Two things people get wrong here:

• The field is just Endpoint. There's only one URL field. Server Transport should already read HTTP Streamable — confirm it does before moving on.
• Use Header Auth, not Bearer Auth. The dropdown offers five options. CNAPS Studio API keys authenticate through a header. If you put your key in Bearer Auth you get 401 invalid token, because Bearer is reserved for OAuth tokens.

Create the credential

Click Set up credential and fill in:

Save. Again — narrow Allowed HTTP Request Domains from its All default. It costs one click and limits where this key can ever be sent.

Step 5 — Select your tools

Setting Tools to Include to Selected reveals a second Tools to Include dropdown. Click it.

If the tool list loads, you're connected. You'll see every CNAPS.ai tool in alphabetical order, from cancel_flow to upload_file. See the MCP Server guide for what each one does.

For this tutorial, check four:

• suggest_flow — recommends which model(s) fit a request
• upload_file — registers an image, video, or audio file
• create_flow — builds the flow
• get_flow — reads status and results

Clear the search box between selections. The filter text stays in the box after you pick a tool. If you type the next tool name without clearing it you end up searching for get_flowupload_file and get "No matching data" — which looks like a connection failure but isn't.

Your node should now look like this:

Step 6 — Test it

Save the workflow. In the Chat panel at the bottom, send:

Upscale this image 4x: https://picsum.photos/id/237/320/240.jpg

Use an image host that allows hotlinking. Some hosts — Wikimedia Commons among them — block non-browser user agents and return HTTP 403, which surfaces as an upload_file failure.

Watch the Logs panel. You should see the agent loop: a model call, then MCP tool calls, then more model calls. Ours made four tool calls and finished in about 40 seconds using roughly 17k tokens.

The reply includes a download URL and a link to open the generated flow in CNAPS Studio:

Input (320x240)

Output (1280x960)

What just happened

The agent wasn't following a script. Given one sentence, it worked out the whole sequence itself:

1. suggest_flow — asked CNAPS Studio which model suits "upscale 4x" and got back PiSA-SR, a super-resolution model

2. upload_file — registered the image URL with CNAPS Studio
3. create_flow — built a three-node flow: Image Loader → PiSA-SR (ratio 4) → Image Viewer
4. get_flow — polled until the run completed, then returned the output URL

Open the Studio link from the reply and you can see, edit, and re-run that flow like any other:

This is the point of the integration. You don't wire up a node per model. You describe the outcome, and the agent assembles the pipeline from the CNAPS Studio catalog.

One caveat before you point the agent at the whole catalog. External models — Gemini, GPT, Claude — run on your own provider key, not on CNAPS Studio's compute. Add the key in CNAPS Studio under your avatar → Workspace Settings → Integrations before asking for a flow that uses one, or the run is rejected before it starts.

Keep Tools to Include on Selected

An AI Agent calls tools based on text it reads. If that text can come from anywhere you don't fully control — a user message, a fetched web page, a file, an email body — then any tool you expose is a tool an attacker can try to trigger.

Setting Tools to Include to All puts every tool in that blast radius, including destructive ones like delete_flow and cancel_flow. Dozens of tool descriptions is a lot of context to weigh for a task that needs four.

Grant only the tools the workflow actually needs. For the workflow in this tutorial, that's four.

Troubleshooting

Server reference

The Tools to Include dropdown always shows the current tool list. For what each tool does, its parameters, OAuth scopes, and example recipes, see the CNAPS.ai MCP Server — User Guide.

Next steps

• Swap suggest_flow for list_models or find_compatible_models if you want the agent choosing from a narrower set
• Add run_batch and get_batch_status to process many files in one run
• Replace the chat trigger with a webhook, schedule, or form trigger to run the same pipeline unattended
• Add a Memory node to the agent so it can refine results across turns ("now sharpen it a bit more")

For the full tool reference and example recipes, see the CNAPS.ai MCP Server — User Guide — or connect the same server to Claude directly.