TL;DR: What n8n v2 really changes
n8n v2 is not a cosmetic update. It's a re-foundation focused on AI orchestration, with notable breaking changes for self-hosted users. Migration is mandatory if you are still running on MySQL or if your workflows use process.env in the Code node.
Key points
- End of MySQL/MariaDB support, PostgreSQL becomes the standard.
- AI Workflow Builder: natural language workflow generation.
- Python task runner replaces Pyodide for executing full code.
- Native AI agents (ReAct, Conversational) and Guardrails node.
- Data Tables: native structured storage without an external database.
- Hardened security: environment variables blocked by default in the Code node.
What exactly is n8n v2?
n8n is an open-source automation platform with over 75,000 stars on GitHub by the end of 2025 (GitHub n8n-io/n8n). Version 2.0 marks the tool's transition to production-grade AI orchestration, with deliberate technical breaks and a focus on agent security.
Specifically, n8n connects over 400 services via visual nodes. v2 retains this nodal logic but adds a layer of intelligence: autonomous agents, prompt-based workflow generation, and full Python execution. The tool is no longer a simple Zapier-like.
Why this v2 now?
The AI automation market has exploded. According to the official n8n blog, the use of AI nodes has increased more than fivefold in one year among cloud users. v1 was not sized for these loads. PostgreSQL becomes the only supported engine to handle multi-agent workflows.
What are the breaking changes in n8n v2?
n8n v2 introduces five major breaking changes documented on the official release notes page. The most impactful: 100% of instances running on MySQL or MariaDB must migrate to PostgreSQL before upgrading. Ignoring these points will break service startup in production.
Security: locked environment variables
In the Code node, access to process.env is now blocked by default. This is a direct response to API key leaks observed on public instances in 2024-2025. To re-enable access, an explicit environment variable must be passed server-side. The ExecuteCommand and LocalFileTrigger nodes are also disabled by default.
End of MySQL and MariaDB
MySQL/MariaDB support is officially dropped. n8n provides a CLI script to migrate SQLite to Postgres, but for MySQL, you need to go through an export/import. The legacy SQLite driver is also removed. This rationalization simplifies maintenance and reduces the bug surface on heavy workflows.
Modified sub-workflow behavior
Sub-workflows containing a Wait Node now return data differently. If you chain sub-workflows with pauses, check your mappings before upgrading. The sub-workflows documentation details the new return format.
How does the AI Workflow Builder work?
The AI Workflow Builder is the flagship feature of n8n v2. You describe your need in French or English, and the tool generates a workflow draft with the correct nodes connected. According to user feedback published on the n8n community forum, prototyping time drops by 60 to 80% for standard cases.
Concrete example: type "retrieve new HubSpot leads, enrich them via Apollo, score them with GPT-4, and push to Slack if score > 8". The builder places the five nodes, configures the credentials placeholders, and draws the connections. You retain control to adjust.
Limits to know
The tool remains in beta. For workflows with more than 15 nodes or complex conditional branches, the AI often proposes a simplified structure that needs manual refinement. For critical workflows, treat the output as a draft, not a final delivery. For automated cold email prospecting, the gain remains massive.
AI Agents and MCP: What do they bring?
n8n v2 natively integrates AI agents according to ReAct and Conversational patterns, popularized by Anthropic via the Model Context Protocol announced in late 2024. An agent can now reason, choose a tool, execute, observe the result, and iterate, without custom scripting.
MCP (Model Context Protocol) support
MCP has become a de facto standard for connecting LLMs to external tools. n8n v2 exposes each workflow as an MCP server usable by Claude, ChatGPT, or any compatible client. The reverse is also true: an n8n agent can consume any third-party MCP server as a tool. The n8n MCP documentation lists available implementations.
The Guardrails node
New in v2, the Guardrails node filters LLM inputs and outputs to block PII leaks, injective prompts, or out-of-scope outputs. Essential whenever your agents touch customer data. Growth teams that industrialize lead scoring via AI gain an almost free layer of compliance.
Why the Python task runner changes the game?
n8n v2 replaces Pyodide with an isolated Python task runner. Pyodide ran in the browser via WebAssembly, with severe limitations on native libraries. The new runner executes full Python in a separate sandbox process, with access to the PyPI ecosystem according to the Code Node documentation.
Concretely, you can finally import pandas, numpy, scikit-learn, or official SDKs like the OpenAI Python SDK. This is a major shift for data teams that used to work around this with HTTP nodes. Caution: existing Pyodide scripts require review, as some global APIs change.
Increased security through isolation
The task runner runs in a separate process from the n8n worker. An exception or crash in your Python code no longer kills the entire instance. For self-hosted production, this is a non-negotiable stability gain.
n8n v2 vs Make, Zapier, and Activepieces: who wins?
n8n v2 positions itself in the segment where Zapier and Make struggle: complex AI workflows, self-hosting, and zero marginal cost at scale. According to n8n's Sustainable Use license, the tool remains free for self-hosted internal use, with no execution cap.
Execution cost
Zapier charges per task, Make per operation. For an AI workflow that chains 10 LLM calls per lead, the cost quickly becomes unreasonable. Self-hosted n8n charges zero for execution; you just pay for your server. For 10,000 leads/month enriched by an agent, the difference amounts to thousands of euros.
Native AI capabilities
Activepieces has a similar approach but lags far behind on advanced agent patterns and MCP integration. Make has released AI modules but without comparable native agent support. Zapier relies on Zapier Central for AI, which is simpler but much less flexible than n8n's LangChain chains.
When to stick with Make or Zapier?
If your team doesn't touch code, your volumes remain low, and you want zero infrastructure to manage, Make remains an excellent choice. n8n v2 still requires some technical aptitude. The ROI shifts in favor of n8n as soon as you process more than 5,000 monthly executions or have critical AI workflows.
What concrete workflows can be built with n8n v2?
n8n v2 opens up growth and sales use cases that v1 made laborious. According to the n8n blog on AI workflow automation, over 60% of workflows published on the template store in 2025 include at least one LLM node. Here are four directly transferable patterns.
Workflow 1: Inbound lead qualification
Trigger on HubSpot form. A ReAct agent retrieves the LinkedIn profile via Apollo, scrapes the website, generates a 0-10 score with GPT-4 according to your ICP, writes a summary in the CRM, and notifies the salesperson on Slack if the score > 7. Build time with the AI Workflow Builder: 20 minutes.
Workflow 2: Daily competitive monitoring
Daily cron. Scraping 10 competitor product pages, semantic diff with Claude to identify pricing or positioning changes, push a Notion digest. The Guardrails node filters hallucinations before publication.
Workflow 3: Editorial content generation
Airtable trigger on new "brief" entry. Multi-step workflow: SERP research via SerpAPI, brief enriched by GPT-4, draft generated, passed through an editorial agent who critiques and rewrites, WordPress publication in draft mode. Relevant for growth marketing teams that industrialize SEO.
Workflow 4: L1 support agent on Gmail
Gmail watcher. Conversational agent that reads the email, searches a Notion database via embeddings, proposes a response to human support via Slack, or responds directly if confidence is high and the client tag is "low-priority". MCP server exposed so the agent maintains a multi-ticket context.
Self-hosted or n8n Cloud: what to choose in 2026?
According to n8n Cloud public pricing, the Starter offer starts at 20 euros/month for 2,500 executions, the Pro plan at 50 euros for 10,000. Beyond 50,000 executions/month, self-hosted almost systematically becomes more cost-effective, provided you have an ops team capable of maintaining Postgres, Redis, and a load balancer.
Cloud: for whom?
n8n Cloud is suitable for teams without DevOps, who want SLA and zero maintenance. The plan includes managed queue mode, automatic backups, and SOC 2 compliance. Ideal for starting or testing without committing to infrastructure.
Self-hosted: for whom?
Self-hosted wins as soon as you have volume, data sovereignty constraints, or custom workflows involving heavy Python. v2 simplifies deployment via Docker Compose or Helm chart, but Postgres remains mandatory. Allow 2-3 days for a clean setup for a production-ready instance.
FAQ
Is n8n v2 compatible with my v1 workflows?
The majority of v1 workflows run without modification. Exceptions concern the Code node using process.env, sub-workflows with Wait Node, and Python scripts on Pyodide. Allow 1 to 4 hours for auditing depending on the complexity of your instance before migration.
How much does n8n v2 cost self-hosted?
The software is free for self-hosted use under the Sustainable Use license for internal use. Your actual costs are limited to infrastructure: a 4 vCPU and 8 GB RAM VPS with Postgres comfortably handles 100,000 executions/month for approximately 30 to 60 euros monthly at OVH or Hetzner.
Does n8n v2 support Claude and GPT-5?
Yes. The OpenAI and Anthropic nodes are up to date with recent models. MCP support also allows connecting any model exposed via an MCP server, including local models served by Ollama or vLLM. The Anthropic credentials doc covers Claude configuration.
Should I migrate immediately?
If you are running on MySQL or MariaDB, yes: support is ending. If you are already on Postgres and your workflows do not use Pyodide or process.env in Code, you can wait 2-3 months to let the Python task runner beta stabilize. Take a snapshot before any upgrade.
Does n8n v2 replace an orchestrator like Airflow?
No, these are complementary tools. Airflow remains more suitable for heavy data pipelines with complex dependencies and fine-grained error recovery. n8n v2 excels in business automation, AI workflows, and multi-SaaS integration. Many teams use both side-by-side.
Should you jump on n8n v2?
n8n v2 is the most ambitious upgrade since the project's inception. The breaking changes are not trivial, but they rationalize the tool for the coming decade. AI Workflow Builder, native agents, MCP, and Python task runner form a combo that repositions n8n against Make and Zapier in the field of production AI workflows.
If your growth or data stack depends on automation, take the time to audit your current workflows, plan the Postgres migration, and test the AI Workflow Builder on a real case. To go further on industrialization, take a look at our Uclic blog on B2B automation projects.
