Docs · Getting started

Before you begin: first steps & requirements

What NIBWP is, what you need, the mental model, and the safety habits to adopt before connecting your first agent.

Before you connect an AI agent to your WordPress site, it is worth spending five minutes understanding what you are about to do — because what you are about to do is genuinely powerful. NIBWP turns your site into a server that AI agents can read from and write to. Used well, that is a superpower: agents that build, fix and maintain your site on command. Used carelessly, it is a foot-gun. This article is the “before knowledge” — the requirements, the mental model, and the safety basics — so that everything you do afterwards is deliberate.

TL;DR

NIBWP exposes WordPress to AI agents as typed tools over a secure endpoint. Before you start you need WordPress over HTTPS, admin access, and an MCP client. Understand three ideas — tools, scopes and the audit log — and adopt three habits — least-privilege tokens, force-draft, and backups — and you will use it safely and confidently.

What NIBWP actually is

NIBWP is a Model Context Protocol (MCP) server that lives inside WordPress. MCP is an open standard for connecting AI applications to tools and data. Instead of an agent scraping your site or you writing custom integration code, NIBWP advertises a set of tools — typed operations like “search posts” or “create product” — that an agent can call over one authenticated HTTPS endpoint. The agent decides what to do; NIBWP validates, authorizes, executes and logs every call.

What you need

Five things. Most you already have:

  • A WordPress site — a recent version, self-hosted or managed.
  • HTTPS — the endpoint must be served over TLS. Cloud-based AI clients (Claude, Cursor) reach your site from the internet, so it must be publicly reachable over HTTPS. Claude Code, running locally, can also reach a localhost or staging site.
  • Admin access — to install NIBWP and generate tokens.
  • An MCP client — Claude Code, Claude Desktop, Cursor, or any tool that speaks MCP. This is the “agent” side.
  • A few minutes — from install to first call is about five.
AI Abilities are powerful — treat them with respectWith the advanced AI Abilities enabled, agents can execute PHP and perform filesystem operations. That is what makes NIBWP capable — and why you should run it on development or staging first, keep backups, and only enable advanced abilities once you understand them. The defaults are safe; you opt into power deliberately.

The mental model: three ideas

Everything in NIBWP rests on three concepts. Understand these and the rest is obvious.

1. Tools (abilities)

A tool is a single typed operation — nibwp/wp-create-post, nibwp/wp-search. Each is self-describing: the agent knows its inputs and outputs without you wiring anything. NIBWP calls these abilities, and ships 30+ of them out of the box, plus more from each active plugin integration.

2. Scopes

A scope is what a token is allowed to do. Tokens carry scopes — read, write, publish, admin — and an agent can only call abilities its token permits. This is your primary safety lever: give an agent the least it needs.

3. The audit log

Every call is recorded — which tool, which token, the arguments, the result, the timing. Nothing an agent does is invisible. The audit log is your accountability and your debugging tool in one.

Three habits that keep you safe

You do not need to be paranoid — you need three habits:

HabitWhy it matters
Least-privilege tokensGive each agent the smallest scope that does the job. A research agent gets read; a content agent gets write, not publish.
Force-draft onMake AI writes land as drafts a human approves. The agent proposes; you dispose.
BackupsA recent backup turns any mistake into an inconvenience. Standard practice, doubly wise with agents.

Your first ten minutes

1

Install on staging

Put NIBWP on a development or staging site first. Activate it to publish the MCP endpoint.

2

Generate a read token

Create an application password scoped to read. Start by letting the agent look, not touch.

3

Connect a client

Use the config generator to connect Claude, Cursor or Claude Code. See the next guides for each.

4

Ask it to read

“List my recent posts.” Watch the call appear in the audit log. You are live — safely.

5

Graduate to write

When you are comfortable, issue a write token with force-draft on, and let the agent draft.

Start by readingThe single best habit is to begin every new agent with a read-only token. You learn what it can see, you build trust, and you risk nothing. Upgrade to write only when you are ready.

What can go wrong (and how it is prevented)

  • An agent changes something unexpected — prevented by scopes and force-draft; everything is a reviewable draft.
  • A runaway loop hammers the database — prevented by per-token rate limits.
  • You cannot tell what happened — prevented by the audit log; every call is recorded.
  • A leaked token — contained by per-client tokens you can revoke individually.

FAQ

Do I need to know how to code?
No — the built-in abilities cover the common cases and you drive everything in plain language. Code is only for adding your own bespoke tools.
Is it safe to run on a live site?
Yes, with the habits above — least-privilege tokens, force-draft, backups. But start on staging to learn the ropes.
Does it work on shared hosting?
Yes, as long as the site is reachable over HTTPS for cloud clients. Claude Code can also reach local installs.
What is the difference between read, write and publish?
Read fetches; write creates and edits drafts; publish makes content live; admin touches settings and users. Grant the least you need.

Ready to begin?

Install NIBWP, generate your first token, and connect a client — the next guides walk you through each.

Get NIBWP

Connect your first agent.

Follow the quickstart, or book a demo and we'll walk you through it.