@datagrout/n8n-nodes-datagrout-mcp
v0.1.0
Published
DataGrout MCP tool sub-node for n8n AI Agents — exposes every server tool as a separate agent tool (self-hosted)
Readme
@datagrout/n8n-nodes-datagrout-mcp
DataGrout MCP — an n8n AI Agent tool sub-node that exposes every tool on your DataGrout MCP server as a separate agent tool (the same experience as the built-in Atlassian/Linear MCP entries). No operation selector: just the credential and a Tools to Include filter.
Self-hosted only. This package ships runtime dependencies (
@langchain/core,zod) and is intentionally not eligible for n8n verification. For a verified, zero-dependency node, see@datagrout/n8n-nodes-datagrout.
What the node handles for you
- Background tasks: slow DataGrout requests are moved to a background task server-side; the node waits and returns the finished result — the agent just receives the answer.
- Lean responses (on by default): large result sets return as a short preview plus a server-side reference that DataGrout's compute tools accept, keeping huge row sets out of your workflow and agent context.
- Connection reuse: calls share one server session instead of re-connecting every time.
Installation (self-hosted)
- Install
@datagrout/n8n-nodes-datagrout-mcpfrom Settings → Community Nodes. - Set
N8N_COMMUNITY_PACKAGES_ALLOW_TOOL_USAGE=trueand restart n8n.
Supported n8n versions: 2.29.x (dependencies are pinned to match; other versions may work but are untested).
Credential — DataGrout API
Same credential as the main package (configure once, use in both): API Token (from your DataGrout dashboard) + Server ID (UUID) + Gateway Base URL.
Usage
- Add DataGrout MCP to an AI Agent's tool connector.
- Pick the credential; choose Tools to Include:
- All — every tool on the server
- Selected — only the chosen tools
- All Except — everything but the chosen tools
- Chat — the agent sees each DataGrout tool separately (name, description, input schema) and calls them directly.
The filter is the permission boundary: filtered-out tools are never exposed to or callable by the agent. With large servers (100+ tools), use Selected to keep the model's context focused.
Click Execute step on the node to preview the tools it will expose.
