Weekend Projects: 3 Practical Builds with MCP, Make AI Agents, and n8n

If you read the tech news this week, one theme keeps showing up: agents are getting “hands”. Not just chat — real tool use, real workflows, and guardrails like sandboxing and standard connectors.

Here are three weekend builds that turn that trend into something you can actually ship. Each one is sized so you can finish it without a venture-backed “innovation lab” — just a laptop and a few focused hours.

Project 1: MCP “Toolbelt” Server for Your Shop Data

Goal: Stand up a small Model Context Protocol (MCP) server that exposes a handful of safe, boring, high-value tools: search customers, create a quote draft, look up parts, and write a follow-up note. MCP is becoming the standard “port” for connecting assistants to tools, so you build the tool once and plug it into multiple hosts later.

Why now: MCP has been getting broad ecosystem attention, and OpenAI’s Agents SDK is explicitly calling out first-class MCP tool support as part of its April 2026 update (AI Automation Global). That’s a signal: build your integrations around MCP and you’ll spend less time rewriting glue code later.

Tools needed: Node or Python, the MCP SDK, SQLite (or a CSV), and one LLM client (anything you already use).
Estimated time: 4–6 hours.
What you’ll learn: Designing small, permissioned tools; how to expose them via MCP; and how to keep “tool use” separate from your business data layer.

Build outline:

Create a tiny data store (SQLite) with customers, jobs, and parts.
Expose 4 MCP tools: search_customers, get_job_history, draft_quote, create_followup.
Add simple safety rules: max rows returned, required filters, and logging of every tool call.
Connect it to an MCP client and run a few realistic prompts (“Quote a water heater swap for Susan, last job was 2024-11”).

Project 2: Make.com “Run an Agent” Intake Bot (With Knowledge + Tools)

Goal: Build an intake assistant for leads and service requests. It asks the right questions, classifies the request (estimate vs emergency vs warranty), and then either schedules a follow-up or creates a clean work order draft.

Why now: Make’s AI Agents (New) app includes a Run an agent module where you can give the agent tools (regular Make modules, scenarios, and even MCP tools) and knowledge files, then test it in a chat sandbox before going live (Make Help Center).

Tools needed: Make.com, your email/CRM module of choice, and one short “policy” document (PDF or DOCX) that explains how you triage leads.
Estimated time: 2–4 hours.
What you’ll learn: Where agents help (messy intake, flexible questioning) vs where classic automation still wins (routing, creating records, sending confirmations).

Build outline:

Create a knowledge file: “Rainshadow Intake Rules” (what you need to quote, what needs a site visit, what is out of scope).
In Make, create an agent and attach tools like Create CRM lead, Send email, and Create calendar event (Make Help Center).
Use a consistent conversation ID per customer so the agent can remember the thread when they reply later (Make Help Center).
Test with 10 real-ish examples (photos, PDFs, short notes). Keep it honest: missing info, blurry details, “call me” messages.

Project 3: n8n Content Approval Workflow (Two-Model “Sanity Check”)

Goal: If you do any outbound — quotes, proposals, follow-ups, newsletters — build a workflow that reviews text before it goes out. One model checks voice and quality; a second model validates and flags disagreements. Then you route it: auto-approve, send to human review, or bounce it back with fixes.

Why now: A strong agent stack still needs checklists and fail-safes. A practical pattern is “AI does the first pass, then a second system verifies it.” An n8n tutorial lays out a concrete version of this idea: webhook trigger, extraction/OCR if needed, primary analysis, secondary validation, then IF/Switch routing and logging (Agile36).

Tools needed: n8n, webhook trigger, one OpenAI node, one Claude (or other model) node, and a Slack/email node.
Estimated time: 3–5 hours.
What you’ll learn: Building reliable decision logic around AI outputs (scores, thresholds, and “send to a human” when it’s fuzzy).

Build outline:

Trigger: webhook or “new draft in folder.”
Optional: OCR or document parsing if the input is a PDF/image (Agile36).
Primary analysis prompt: score voice compliance, find risky claims, suggest edits, return JSON (Agile36).
Secondary validation: re-check and flag discrepancies (Agile36).
Switch logic: auto-approve if both scores are high and no legal flags; otherwise route to review (Agile36).

Pick one and ship it

If you only do one thing this weekend, do Project 1. A clean MCP toolbelt forces you to define your “shop brain” as a set of safe tools — and that’s the hard part. Once it exists, you can plug it into whatever agent runner wins next quarter.

If you want the fastest visible win for a small business, do Project 2. Lead intake is messy, and agents shine when the inputs are messy.

And if your biggest pain is “we sent something that sounded wrong,” do Project 3. A boring approval gate saves more reputation than any fancy prompt.