How can I bypass the token limit issue?

[ArtemSultanov-PG]

Hello everyone. If the model is running out of tokens, how can I set up the workflow so that:

1. The cursor restarts the request to the model after it fails.
2. The model receives the status of the last operation, rather than starting the entire operation cycle from the last task sent to the agent.
3. The model summarizes all the executions and stores them in supabase/

[czlonkowski]

Hey @ArtemSultanov-PG — so the question splits into two parts that have different answers:

Part 1: cursor / model session management. "Restart the request after failure, resume from last operation, summarize and store in supabase" — that is not really an n8n-mcp thing, that is how you orchestrate the LLM client. In Cursor / Claude Code, your options are:

• Use long-running compaction (Claude Code does this automatically, condensing old turns into a summary).
• Use a sub-agent / task tool for the heavy multi-step part so the parent context stays small.
• Persist intermediate state yourself — write the partial workflow JSON to disk between turns, so a fresh session can pick up where the last one died.

If you specifically want supabase as the durable store, you can build that loop in n8n itself with the same n8n-mcp exposed via HTTP — the agent calls Supabase tools to log progress, then calls n8n-mcp to apply the next step.

Part 2: keep n8n-mcp from chewing through tokens in the first place. This is where I can actually help. Three things make a huge difference:

1. Use get_node at detail: "standard" (the default) — that is ~1-2k tokens vs ~3-8k for full. Only escalate to full when you genuinely need every property. Do not use full as a habit.
2. Use n8n_update_partial_workflow instead of n8n_update_full_workflow — it sends operation diffs (addNode, removeConnection, patchNodeField, etc.) instead of the whole workflow on every edit. Saves 80-90% on update turns. There are 18 supported operations covering nodes, connections, settings, tags, and activation.
3. Install the n8n-skills pack (/plugin install czlonkowski/n8n-skills in Claude Code). Skills load on demand instead of sitting in your system prompt — so the model only pulls in expression syntax / validation patterns / etc. when it actually needs them, instead of paying the token cost on every turn.

Try those three together and you should see your effective context budget for an n8n session roughly triple. If you hit a specific case where the model is burning context on something dumb, paste a transcript and I will tell you which knob it is missing.

[czlonkowski]

Also tracking the broader code-execution-with-MCP exploration in #716 if you want to follow along.