AI agent + MCP server: A production AI agent with tool use and guardrails, plus an MCP ser

LangChainAnthropic APIOpenAI APIMCPQdrantPythonTypeScriptAWS

01The problem

Support and ops work meant searching docs and internal tools by hand. An LLM could do a lot of that, as long as it could reach real data without being able to break anything.

02Approach

I built an agent on the Anthropic and OpenAI APIs with LangChain. It plans, calls tools, keeps short-term memory and runs every action through guardrails.

The tools live in an MCP server. Read tools answer questions. Action tools need a human to approve them, and anything destructive isn't exposed at all.

AI with guardrails

Tool use is scoped, calls are audited, and eval harnesses catch regressions before users do.

03Retrieval and evals

Answers that need documentation go through retrieval over embeddings in Qdrant. An eval harness runs on every change so a prompt or model update can't quietly make answers worse.

04Results

The agent and the tool server run in production. The demo below is a scripted trace that shows the same pattern with sample data.

Live artifactTry it yourself

Runs in your browser. Nothing to install.

Interactive demo — embedded on xerk.io. The article above explains what it shows.

MCP

Server in production

RAG

With eval harnesses

Model providersAnthropic + OpenAI

FAQ

What is an MCP server?

A Model Context Protocol server exposes tools to LLM clients in a standard way, so an agent can use internal systems through controlled calls you can audit.

Can you add an AI agent to my product?

Yes. Tool use into your APIs, retrieval over your data, guardrails and evals. See the AI agent service on the hire page.

Want an AI agent like this?

I'm Ahmed Mamdouh — 10+ years building systems like AI agent + MCP server. Scoped in one call.

Start a mission Hire on Upwork