( Services )
AI Agents and MCP Servers for Business
I connect Claude, ChatGPT and other models to your WordPress site, store, CRM or ad accounts, so they can do the work instead of just talking about it. A human signs off on every step that matters.
An AI agent is a language model that carries out tasks in your systems (reading data, editing content, preparing reports) instead of only answering questions. An MCP server is how you give it that access safely. I build both, starting at $2,200, always with a human approving the steps that matter.
What I run myself
I use AI agents every day in my own business. Everything I offer, I have tested on myself first:
- Blog agent. On my server, a model drafts a post, picks an image and emails it to me. I click "Publish" or I don't. Without that click, nothing goes live.
- MCP server for WordPress. I wrote a plugin that exposes around 20 content operations: posts, Yoast meta, ACF fields, media, categories, templates. The agent fixes pages, writes SEO descriptions and tidies up tags, and I can see every change it makes.
- Ad data agent. It reads Google Ads and Meta Ads campaigns through their APIs, and I ask in plain language: "which ads are burning budget?" It is read-only and never changes anything in the campaigns.
- Server agent. An agent on my server takes technical tasks I send it from my phone. It has written rules about what it must not do.
When an agent makes sense
It makes sense when the task needs reading and judgement, not just moving data around. Rewriting product descriptions to a set of guidelines, reviewing ad campaigns, drafting a reply to an unusual enquiry, cleaning up content on a website.
It doesn't when the steps are always the same. Then ordinary automation in n8n will do the job cheaper, faster and more predictably. An agent is slower and costs more to run. It pays off where it replaces thinking, not clicking.
Security comes first
An agent with access to your store or website is a real risk if it is built carelessly. So:
- Separate account, minimum permissions. The agent gets exactly the operations it needs. My WordPress MCP server checks permissions on every single operation, not just at login.
- Reversible, not destructive. Instead of deleting, the agent hides things or flags them for review.
- Human approval. Publishing, sending and price changes wait for your yes.
- Logging. Every operation is recorded, so you know what the agent did and when.
Which models it works with
MCP is an open standard, so the same server connects to Claude, ChatGPT and any other tool that supports it. You are not tied to one vendor: if a different model is better a year from now, you switch the model, not the whole setup.
If you need a full customer-facing app with AI rather than an agent, see custom AI applications. For simpler, rule-based workflows, start with AI automation.
How an agent differs from a chatbot
A chatbot answers
It gets a question, looks through the knowledge you gave it and replies. It changes nothing in your systems.
An agent acts
It gets a task and picks its own tools: reads data, edits a post, sets SEO meta, drafts a report. It can do exactly what your permissions allow, and nothing more.
A human approves
Anything that publishes, sends or deletes waits for your sign-off. The agent prepares the work, you make the call.
MCP server for WordPress
My plugin exposes around 20 operations: posts, Yoast meta, ACF fields, media, categories, templates. The agent edits the site without logging into wp-admin.
Ad data agent
Ask in plain English what is happening in your Google and Meta campaigns. The agent reads the data through their APIs and answers with numbers. Read-only.
Content agent
Drafts a blog post, picks an image and emails you a "Publish" button. I run one on my own site every day.
How I build an AI agent
-
01
The task and its limits
We agree what the agent should do, what it can access and which steps need your approval.
-
02
MCP server
I expose your system as a set of safe operations: a separate account, minimum permissions, everything logged.
-
03
Agent and testing
I connect the model and test it on a copy or in read-only mode before it touches production.
-
04
Day-to-day use
The agent runs from your laptop, from your phone or on a schedule on a server.
How much does an AI agent cost?
MCP server for your system
from $2,200
- WordPress, online store, CRM or your own app
- Read and write operations with permission checks
- Works with Claude, ChatGPT and other MCP clients
- Setup guide for your team
AI agent for a specific task
from $3,300
- Agent with tools and human approval
- Run on demand or on a schedule
- Notifications by email or to your phone
- AI model usage billed separately
Tools I use
- Model Context Protocol (MCP)
- Claude Code
- Anthropic Claude
- OpenAI
- WordPress Abilities API
- WordPress REST API
- Google Ads API
- Meta Marketing API
- n8n
- Python
- PHP
Frequently asked questions
An AI agent is a program built on a language model that carries out tasks instead of only answering questions. It decides which tools to use (read data, change something, draft a document) and in what order. What separates it from a chatbot is access: it can act inside your systems, within the permissions you give it.
MCP (Model Context Protocol) is an open standard for connecting AI models such as Claude and ChatGPT to outside systems. An MCP server sits on your side and tells the model: these are the operations you may run, with these parameters. The same server works with different models, and you stay in control of exactly what the agent is allowed to do.
With me, an MCP server for one system, such as a WordPress site or an online store, starts at $2,200. An agent built for a specific task, with tools, scheduling and human approval, starts at $3,300. On top of that you pay for model usage; for a typical agent that runs a few times a day, that is usually somewhere between $10 and $150 a month.
It can if it gets too much access, which is why it does not. The agent works from a separate account with minimum permissions, delete operations are either switched off or reversible, and anything it publishes or sends waits for your approval. Every action is logged, so you can see what it did and when.
If the process has fixed steps (an invoice arrives, copy it into your accounting system), plain automation is enough, and it is cheaper and predictable. An agent makes sense when the steps depend on the situation: something has to be read, judged and a decision made about what comes next. I often combine the two, with n8n starting the agent on a schedule and the result going back to a person for approval.