n8n 2.35 Beefs Up Local AI Agents — Selfhosted Automation Is Now a Serious Agentic Platform

On August 11, 2026, n8n shipped version 2.35 of its workflow automation platform. The changelog is modest — an AI agent context leak fix, a redaction policy for Code node output, better tool result bounding — but the signal is elsewhere. By mid-2026, a selfhosted workflow engine can run AI agents connected to hundreds of APIs, with secret isolation, native MCP (Model Context Protocol) support, and an audit console.

What was a hobbyist sandbox in 2024 has become a production-grade agentic platform that enterprises deploy behind their firewall in place of Zapier or Make.

The real feature in n8n 2.35: AI agents learn to shut up

Three fixes in this release directly target the AI Agent nodes:

1. Pre-tool-call text leak prevention (#35212). n8n’s AI agents, when calling a tool, would sometimes generate residual text before invocation that « leaked » into the final response. The fix cleans this noise — a technical detail, but an agent that babbles before acting is an agent whose logs are useless in production.
2. Tool result bounding (#35759). Oversized tool results (e.g., a SELECT * without LIMIT) were passed in full to the LLM, saturating the context window and generating unnecessary API costs. Version 2.35 truncates these results to a configurable size.
3. MCP schema compilation (#35610). Tools exposed through the MCP protocol are now compiled against their declared JSON Schema dialect, improving compatibility with third-party MCP servers.

These three fixes share the same DNA: they graduate n8n’s AI agents from proof-of-concept to reliable tooling. It’s the difference between a workflow that works « most of the time » and one you can trigger from a critical webhook without human supervision.

Secret redaction in Code nodes: closing a blind spot

Fix #35121 may be the most important one for security teams. n8n’s Code nodes — the equivalent of an arbitrary eval() in JavaScript or Python — let you write custom business logic. Before version 2.35, console.log() inside these nodes could accidentally print secrets (API keys, tokens, passwords) to the execution console.

The new redaction policy automatically applies a filter to all console.log() output in Code nodes. Strings matching credentials stored in n8n’s secret manager are replaced with [REDACTED].

This is a mechanism GitHub Actions implemented in 2023 (the famous ***) and GitLab CI followed in 2024. n8n catches up to the industry standard — and for a selfhosted platform, this is a significant maturity milestone.

The landscape: n8n vs. Activepieces, Node-RED, and cloud iPaaS

By August 2026, the selfhosted automation market has coalesced around four categories:

n8n occupies a unique position. It’s the only tool that combines native AI agents, selfhosted execution (Docker / Kubernetes), and a Sustainable Use License (source-available, not OSI open source). This license prohibits SaaS resale but allows any internal deployment — a tradeoff enterprises accept because they’re not reselling their internal n8n instance.

Activepieces, the direct challenger, is MIT-licensed and offers a cleaner UX, but its AI agent and MCP support lags six to twelve months behind n8n. Node-RED remains unbeatable for IoT workflows and lightweight hardware (Raspberry Pi), but its visual programming model doesn’t scale well to complex agentic workflows.

Why enterprises are ditching cloud automation for selfhosted

The « data sovereignty » argument for selfhosting is tired. The real reasons for the shift in 2026 are economic and technical.

Cloud iPaaS costs explode with volume. A Zapier workflow handling 50,000 tasks per month costs $133/month on the Professional plan — and that’s per workflow. A team of ten people with five workflows each quickly hits $6,000/year, excluding overages. A selfhosted n8n instance on a $6/month VPS handles the same volume with zero marginal cost.

Agentic workflows consume LLM tokens, not « Zapier tasks ». A GPT-5.6 call with a 32K-token context costs roughly $0.06. A workflow that calls an LLM fifty times a day generates $90/month in API costs — on top of the Zapier/Make subscription. With selfhosted n8n, you pay only for tokens, not the platform.

Latency matters. A workflow orchestrating an AI agent, a SQL query, and an outgoing webhook has an end-to-end latency of 200 ms on selfhosted n8n (everything is local). On Zapier, each step routes through Zapier’s servers, adding 500 ms to 2 s per node.

Verdict

If you run fewer than 10 simple workflows and don’t want to manage infrastructure, stick with Zapier or Make. The comfort of not administering a server is worth the subscription.

If you run agentic workflows (LLMs, MCP, RAG) or your volume exceeds 50,000 executions per month, deploy n8n selfhosted now. Version 2.35 proves the platform is ready for agentic production — and the savings pay for the server in two months.

If you’re in IoT or lightweight prototyping, Node-RED remains the best choice. But for anything involving an LLM, the game is over.

Selfhosted automation is no longer the « weekend hacker’s » choice. By August 2026, it’s the rational choice for the team that does the math.

References

• n8n Release Notes, « [email protected] », August 11, 2026
• GitHub, n8n-io/n8n, releases/tag/[email protected]
• Activepieces Documentation, « Self-hosting Activepieces », accessed August 11, 2026
• Zapier Pricing, « Plans and Features », accessed August 11, 2026
• Model Context Protocol, « Specification », modelcontextprotocol.io, 2026