The local setup, free and yours
Yesterday used the n8n cloud trial. Today it runs on your own machine, which costs nothing and has no token limit.
node --version # anything above 13 is fine
npm --version # version does not matter
npm i n8n -g # same command on Windows and Mac
npx n8n # prints a local URL, open it
You create an account on first run, and it is local, nothing leaves your machine. If you already had one and forgot the password, there is a user-management reset command rather than a reinstall.
For the brain, Ollama with a small model. Only the URL is needed; the API key field is optional.
| Model | Works on |
|---|---|
| Gemma 3 1B | 8GB RAM and up |
| Llama 3.2 3B | 8GB RAM and up |
The one that bites: not every local model can call tools. Gemma answered chat fine but could not drive the Jira agent, because it does not support tool calling. Switching to Llama 3.2 3B fixed it, and the agent then fetched a real ticket and generated a test plan. If your local agent silently refuses to use its tools, that is the first thing to check, and it is a model capability rather than a wiring mistake.
And the standing rule, repeated again: do not install this on a company laptop without permission.
The six agents
All six are committed in chapter_08_n8n/Agents/ as importable JSON:
| # | Agent | Does |
|---|---|---|
| 1 | Fetch Jira ticket | Reads a ticket by key |
| 2 | Create Jira ticket | Files a bug from a description |
| 3 | Fetch Jira and create test cases, local LLM | The same job on Ollama, no cloud |
| 4 | Bug triage | The one below, and the one that matters |
| 5 | RCA automation | Root cause on a production bug |
| 6 | Social media | Schedule, write, illustrate, post |
The bug triage agent
The shape is three nodes, and the whole thing is readable at a glance.
Two prompt files are committed rather than one, and the pair is the lesson: 01_Raw_BugTriage.prompt.md is the first draft, and 02_ModifiedBugTriagePrompt.prompt.md is the same intent rewritten to be tool-aware after ChatGPT was asked to improve it. Keeping both lets you see what changed and why, which is the same instinct as the BLAST tracking files from yesterday.
This agent is at roughly twenty per cent on its own, and the session said so plainly rather than overselling it. It has no company context: it has never seen your Confluence, your PRDs, or your test repositories. Adding RAG is what takes it to around seventy. The last thirty per cent is human, permanently. The pitch is not that review disappears, it is that the review meeting shrinks: the same method as yesterday's triage crew, where the arithmetic is worked through properly.
The RCA agent, and a warning about building with AI
The RCA agent was built by asking ChatGPT for both the prompt and the n8n workflow JSON, then importing the JSON straight into n8n. It works, and the session immediately put a fence around it:
Do not use n8n's build-with-AI or JSON import until you have built four or five workflows by hand. The reason is the same one from the hiring conversation: importing a workflow you cannot read produces something you cannot debug, cannot explain, and cannot modify when it breaks. Build them manually first, then let AI accelerate what you already understand.
The social media agent, and an honest number
The sixth agent is a content pipeline: a 9am schedule trigger, pick a topic, draft a LinkedIn post, generate an image with Gemini, add hashtags, merge, review, post. It can be one agent with Gemini attached as a tool, or several smaller agents, and the modular version is easier to tune because you can fix one step without disturbing the rest.
Then the part worth repeating, because most people selling this will not tell you:
Pure AI posts underperform badly. The figures given from real accounts: AI-written posts land somewhere around seven to thirteen likes, against roughly two hundred and seventy to a thousand for posts written by hand. The conclusion drawn was not "do not automate", it was mix the two: automate the drafting, the image and the scheduling, and keep a human in the part that makes it sound like a person. Worth holding next to the 60-day branding challenge.
For distribution, an upload-post.com node can fan one post out to nineteen or more platforms, with a free tier of ten a month.
Tasks and announcements
- Replicate the five agents built in class, and post screenshots or proof of them running.
- Build the sixth, the social media agent, and build the RCA agent from the shared prompt.
- Code, prompts and the sample spreadsheet are in
chapter_08_n8nof the AITesterBlueprint4x repository. - Coming next: Langflow, RAG, MCP, CrewAI, Python and LangChain.
- Certification sessions are being scheduled this week, targeting AI Fluency part 2 and the Claude architect certification.
- Previous: BLAST and the first agent that runs without you. Related guides: n8n for QA and 10 n8n QA workflows.