Reducing token usage?

[skyhi69]

I apologize as a older person, I am just learning to be more technical and struggling to get caught up on the AI revolution by learning n8n, and learning to use this wonderful program to help me create workflows. I read that the biggest complaint seems to be about token usage limits with Claude using tools and mcp, and I am wondering if there are ways to mitigate this, with the latest research by Anthropic (www.anthropic.com/engineering/code-execution-with-mcp)? Is it possible to optimize according to these findings?

[czlonkowski]

Hey @skyhi69 — no apology needed, this is exactly the right question to ask.

Most of what the Anthropic post is pointing at is "stop pulling huge MCP responses into the context window when you do not have to". The good news is n8n-mcp is already built around that idea. Here is what is in your toolbox today:

• get_node has three detail levels: minimal (~200 tokens), standard (~1-2k, the default), full (~3-8k). Default to standard. Only use full for nodes where standard is genuinely missing what you need. People burn enormous amounts of context by reflexively asking for full on every node.
• n8n_update_partial_workflow sends operation diffs instead of the whole workflow JSON on every edit — addNode, patchNodeField, addConnection, and 15 others. The project measures around 80-90% token reduction on update turns vs the full-workflow update path. If you are iterating on a workflow, this is the single biggest win.
• The n8n-skills companion pack (/plugin install czlonkowski/n8n-skills) ships seven Claude Code skills — expression syntax, validation patterns, node config, JS/Python in Code nodes, MCP tool usage, workflow patterns. Skills load on demand, not in the system prompt, which is exactly the "do not pay for what you do not use" pattern Anthropics post is advocating. If you are not running these yet, install them — biggest single quality-of-life upgrade I can recommend.

On the Anthropic "code execution with MCP" pattern specifically — that is the stronger version where the MCP server exposes a sandboxed code-execution surface and the model writes scripts that filter/transform tool output before any of it ever reaches the context window. n8n-mcp does not do that today. It is a real pattern and a sensible direction, but it is a meaningful build. I will open an issue to track exploring it as a future direction. For now the three knobs above will get you most of the way there.

Welcome, by the way — you are asking the right kind of questions. You will be fine.

[czlonkowski]

Tracked as #716 — happy to discuss the design over there.