Build a Daily AI Newsletter Agent in n8n
For Solo operators and creators · Based on Alejandro AO Agentic RAG n8n Build
// TL;DR
Solo operators can build a daily newsletter agent in n8n that researches current events and formats a briefing — with no ingestion workflow or vector store required. You use a single Retrieval Workflow: a Schedule trigger (daily) → AI Agent with only the Firecrawl MCP tool → a system prompt instructing it to research your topics and produce a formatted newsletter → an output node that emails or Slacks the result. Running on cheap open LLMs like MiniMax M2.7 keeps costs near zero. Because all content is live-fetched, you skip embeddings entirely.
Do you need a vector store for a research agent?
No. If your agent only needs live, up-to-date information — current events, trending topics, competitor news — you skip the entire ingestion pipeline and vector store. There are no documents to embed. Instead, you give the agent live web capabilities and let it research on demand. This is the simplest form of Agentic RAG: an agent with a web-search tool and nothing else.
That makes it ideal for solo operators who want a hands-off daily briefing without managing document ingestion or a persistent database.
How do you build the daily newsletter workflow?
Create one Retrieval Workflow. Start with a Schedule trigger set to fire once daily at your preferred time. Connect it to an AI Agent node.
Attach an OpenAI Chat Model node routed to router.huggingface.co/v1 running a cheap, capable open model — MiniMax M2.7 or Kimi K2.6 both handle agentic research well at $1–$3 per million tokens, versus $30+ for closed frontier models. Use your Hugging Face token with 'call inference providers' permission enabled.
Add an MCP Client Tool node set to 'Streamable HTTP' pointing at the Firecrawl MCP endpoint (include your Firecrawl API key in the URL). The agent auto-discovers Firecrawl's search, scrape, and crawl tools.
How do you get a well-formatted briefing?
The system prompt does the heavy lifting. Write something like: 'Each day, research the latest developments on [your topics] using Firecrawl. Summarise the 5 most important items as a formatted newsletter with headlines and 2-sentence summaries. Do not scrape entire large pages like Wikipedia — you will exceed context.' That last heuristic matters: letting the agent scrape huge pages exhausts the context window and can fail the run.
Finally, add an output node — Gmail to email the briefing to yourself, or a Slack node to post it to a channel.
How do you make sure it actually works?
Before relying on the schedule, test manually. Trigger the workflow and check the execution logs to confirm the agent actually calls the Firecrawl MCP tools and returns formatted output. Inspect tool call counts and watch for thinking loops. Run the complete workflow end-to-end rather than testing nodes in isolation, since data-passing bugs between the agent and the output node only surface in a full run.
Monitor the Executions panel over the first week to track failure rate and refine your system prompt.
What can you expect from the finished agent?
A fully automated daily email or Slack briefing researched live from the web, running at negligible cost thanks to free embeddings being unnecessary and cheap open LLMs powering the research. No database to maintain, no documents to manage — just a scheduled agent that reads the internet and writes you a briefing.
Next step: Spin up n8n (locally with `npx n8n` to start, or Docker on a VPS for always-on scheduling), get a Firecrawl API key and a Hugging Face token, and build the single Schedule → AI Agent → Firecrawl → Gmail workflow. Test it manually before enabling the daily schedule.
// FREQUENTLY ASKED QUESTIONS
Do I need any documents or a vector store for this?
No. A daily research newsletter agent fetches all content live via Firecrawl MCP, so there's nothing to ingest or embed. You skip the Ingestion Workflow and vector store entirely. You only need a Schedule trigger, an AI Agent with the Firecrawl MCP tool, and an output node like Gmail or Slack.
How do I stop the agent from failing on huge web pages?
Add an explicit heuristic in the system prompt: 'Do not scrape entire large pages like Wikipedia — you will exceed context.' Scraping massive pages exhausts the LLM's context window and can cause the run to fail or truncate. Guiding the agent to use search results and targeted scrapes keeps runs reliable.
How much does a daily newsletter agent cost to run?
Very little. Using open models like MiniMax M2.7 or Kimi K2.6 via Hugging Face inference providers costs roughly $1–$3 per million output tokens — a fraction of the $30+ closed frontier models charge. With one run per day and no embeddings needed, your main variable cost is Firecrawl usage for web research.