Custom Node as an AI tool

[ihor-sokoliuk]

Bug Description

I am working on a custom node that an AI tool would need to use. It’s a SearXNG tool:
https://www.npmjs.com/package/n8n-nodes-searxng
Analog to SerpApi. It performs a search request.

I can see the node in the list of AI tools after adding a usableAsTool: true to the node source code (like in NocoDB standard tool).

I can add the NocoDB as a tool and it works perfectly fine, but not the custom SearXNG node.

Once I run a test on the workflow, it fails after the trigger with the error:

Unrecognized node type: n8n-nodes-searxng.searxngTool
Error: Unrecognized node type: n8n-nodes-searxng.searxngTool
    at LazyPackageDirectoryLoader.getNode (/usr/local/lib/node_modules/n8n/node_modules/n8n-core/dist/DirectoryLoader.js:160:15)
    at LoadNodesAndCredentials.getNode (/usr/local/lib/node_modules/n8n/dist/load-nodes-and-credentials.js:272:23)
    at NodeTypes.getByNameAndVersion (/usr/local/lib/node_modules/n8n/dist/node-types.js:36:51)
    at new Workflow (/usr/local/lib/node_modules/n8n/node_modules/n8n-workflow/dist/Workflow.js:56:39)
    at TestWebhooks.toWorkflow (/usr/local/lib/node_modules/n8n/dist/webhooks/test-webhooks.js:290:16)
    at TestWebhooks.needsWebhook (/usr/local/lib/node_modules/n8n/dist/webhooks/test-webhooks.js:169:31)
    at WorkflowExecutionService.executeManually (/usr/local/lib/node_modules/n8n/dist/workflows/workflow-execution.service.js:114:58)
    at WorkflowsController.runManually (/usr/local/lib/node_modules/n8n/dist/workflows/workflows.controller.js:293:16)
    at handler (/usr/local/lib/node_modules/n8n/dist/decorators/controller.registry.js:93:24)
    at /usr/local/lib/node_modules/n8n/dist/response-helper.js:110:26

To Reproduce

1. Install community node n8n-nodes-searxng over n8n Web UI
2. Add AI Agent (Tool Agent)
3. Click on Tools plus button
4. Find and add SearXNG tool
5. Define credentials for it
6. Run the workflow
   

Expected behavior

AI Agent node works and calls the SearXNG tool.

Operating System

Official Docker Image (n8nio/n8n:next)

n8n Version

1.74.1

Node.js Version

20.18.0

Database

SQLite (default)

Execution mode

main (default)

[Joffcom]

Hey @ihor-sokoliuk,

We have created an internal ticket to look into this which we will be tracking as "N8N-8101"

[netroy]

This is intentional. We have a check here to only allow core nodes as Tools for now, until we are certain that the feature is mature enough to allow custom nodes as well.

Considering that a lot of the AI related interfaces in the engine aren't stable enough yet, If we start allowing custom nodes as tool now, we'd have to maintain backward compatibility for all those nodes.

[ihor-sokoliuk]

@netroy Thank you for your prompt response!

Can I track the progress somewhere? So that I would know when I can try this new feature

[netroy]

We are in the middle of re-architecting some of the AI nodes support code, but there is no clear ETA that I can share with you.
However the good news is that this work is also blocking some important upcoming features, so hopefully we can share an update with you and the community soon.

In the meanwhile, if you want to still use the SearXNG node as tool, you could fork the repo, remove the check for I linked above, and build a custom image from that.
Let me know if you would like to try that option, until start supporting tools support for custom nodes.

[netroy]

I've triggered a custom docker build on my fork here.
Once the job finishes, you should be able to use ghcr.io/netroy/n8n:custom-nodes-as-tools as the docker image.

[ihor-sokoliuk]

@netroy Wow! It works like a charm!

I planned to fork and adjust the logic but you made it first.

I will keep using this docker image.

Thank you so much!

[hyroescom]

@ihor-sokoliuk Is that also allow to create custom nodes as Model? I want to create custom model that will enable prompt caching for Anthropic AI. It's super easy but I can't create node for that on current n8n.

[netroy]

@hyroescom This issue is about using existing nodes as tools on AI agents. What you are asking for is a custom model nodes. These are completely unrelated things.
You should already be able to create custom models, since there are no checks in code to clock those.
However if that does not work, please open a feature request on community.n8n.io instead.

[netroy]

@ihor-sokoliuk just FYI, I've also been testing a simple wrapper over the existing SearXNG tool from langchain for over a month now. Just opened a PR, in case you want to give it a try.