Connect AI agents to your content workflow with MCP

The Model Context Protocol (MCP) is an open standard that lets an AI assistant like ChatGPT, Claude, Grok or Cursor use tools in other apps. Connecting your content system over MCP lets the assistant plan topics, commission drafts and follow up on reviews from a chat. The safe setup lets agents plan and draft while a person approves.

This guide explains what MCP is, what an assistant can usefully do with your content workflow, how to keep it safe, and how to connect each of the four most common assistants.

What MCP is

MCP’s official documentation defines it as “an open-source standard for connecting AI applications to external systems.” It compares MCP to a USB-C port: one standard plug instead of a custom cable for every device.

Anthropic introduced MCP in November 2024 as an open standard for two-way connections between data sources and AI tools. MCP’s documentation now lists support in ChatGPT, Claude, Cursor, Visual Studio Code and many other clients.

Three terms are worth knowing:

• Server. The app that offers tools, such as your content system. It describes each tool: its name, what it does and what inputs it takes.
• Client. The assistant that connects to the server and lets the model call those tools.
• Tool. One action, such as “list my workspaces,” “commission an article” or “get the review notes for this draft.”

A remote MCP server runs on the web at a URL. You give the assistant that URL, sign in once, and the assistant can call the server’s tools whenever a conversation needs them.

Why connect an assistant to your content system

Most content work happens in a few apps: a planning doc, a writing tool, a review thread, a publishing platform. An assistant connected to all of it can do the connecting work for you.

Useful jobs for a connected assistant:

• Planning. Read your strategy and search data, suggest topics, and add ideas to the calendar.
• Commissioning. Start a draft with the right title, keyword and facts, without filling in a form.
• Following up. Check which drafts are waiting for review, who is assigned, and what the reviewer said.
• Revising. Turn a reviewer’s comments into a revision request.
• Reporting. Summarize what was published this month and how pages are doing in search.
• Publishing approved work. Send a version a person already approved to your site, or schedule it.

The gain is less about speed and more about context. The assistant you already use for research can act on what it finds, in the system where the work lives.

The safety model: agents draft and plan, a person approves

Connecting an AI to tools that change things deserves care. The MCP specification itself says, in its tools specification, that “there SHOULD always be a human in the loop with the ability to deny tool invocations.”

The vendors say the same. OpenAI’s MCP documentation warns that instructions hidden in content, such as a web page, can cause a model to take unintended actions, and that a server can mislabel a write action as read-only. Anthropic’s guide to custom connectors advises: “Only connect Claude to servers built and hosted by organizations and applications you trust.”

For a content workflow, those warnings point to a clear split:

Action	Agent over MCP	Person in the app
Read plans, drafts, review notes and search data	Yes	Yes
Plan topics and edit the calendar	Yes	Yes
Commission drafts and request revisions	Yes	Yes
Comment and assign a reviewer	Yes	Yes
Approve a draft	No	Yes
Publish an approved version to a website	Yes	Yes
Post to social networks or email a newsletter	No	Yes
Withdraw a live page	No	Yes
Change team access, billing or destination credentials	No	Yes

The logic is simple. Agents may do anything that produces a draft or a plan, because a person reviews it next. Approval is the checkpoint, so it stays in the web app with a signed-in person. Actions that can’t be taken back, like a social post or a sent email, also stay with a person.

Three more safeguards matter:

1. Scoped access. Sign-in grants read access, write access, or both. Write tools ask for more only when needed.
2. Your role still applies. An agent acts as you. If you are a viewer, it can’t write, whatever it asks for.
3. Agent work is marked. What an agent writes or changes is recorded as the agent’s, so you can see it in the activity log and treat it differently.

Our guide to a content approval workflow for small teams covers who approves what in more detail.

How sign-in works

Remote MCP servers usually use OAuth, the same sign-in pattern as “Sign in with Google.” The MCP authorization specification builds on OAuth 2.1 and requires PKCE, which protects the sign-in code from being intercepted.

In practice, you paste the server URL into the assistant, it opens a sign-in page, you choose what access to grant, and you click Allow. There is no password or key to copy. The connection lasts 30 days, then you authorize it again, and you can disconnect at any time.

Some setups can’t run a browser sign-in, such as a bot built on a model API. Those connect with an API key sent in a header instead. Keep that key in a secret store, never in a committed file.

Example prompts that work well

Once connected, talk to the assistant as you would to a colleague who has access to your content tools. These requests work well:

• “List my workspaces and tell me which drafts are waiting for review.”
• “Look at the last 28 days of search data and suggest five topics we rank for but don’t have a page on.”
• “Add three posts on spring lawn care to the calendar for April, one a week.”
• “Commission an article titled ‘How to winterize a sprinkler system’ for the Denver site, targeting ‘sprinkler winterization Denver’.”
• “What did the reviewer say about the pricing page draft? Request a revision that fixes those points.”
• “Which approved articles aren’t published yet? Schedule them for their calendar dates.”
• “Summarize what we published this month and how each page is doing.”

Be specific about the site, the date and the outcome you want. A well-built server offers no approval tool, so an assistant asked to approve a draft should send you the review link instead.

How to connect ChatGPT, Claude, Grok and Cursor

The steps below use Wavefront’s server URL, https://wavefront.adwave.com/mcp, as the example. Any remote MCP server that supports OAuth connects the same way. Menu names in these apps change often, so check each app’s own docs if a label has moved.

ChatGPT

ChatGPT adds custom MCP servers from its settings, and OpenAI’s MCP documentation describes the current flow. Paid plans may need developer mode turned on first.

1. Add a custom MCP server and enter the server URL.
2. Choose OAuth as the authentication method.
3. Sign in on the consent page that opens, keep the requested access and click Allow.
4. In a chat, add the connector from the tools menu and ask it to list your workspaces.

Our ChatGPT integration page keeps the current menu path.

Claude

Claude supports custom connectors on its web and desktop apps. On Team and Enterprise plans, an owner adds the connector for the organization first, then each member connects.

1. Open Claude’s connector settings and choose Add custom connector.
2. Name it, enter the server URL, and leave the client ID and secret empty.
3. Click Connect, sign in, keep read and write access checked, and click Allow.
4. In a chat, turn on the connector and ask Claude to list your workspaces.

In Claude Code, run claude mcp add --transport http wavefront https://wavefront.adwave.com/mcp, then /mcp to sign in. The Claude integration page keeps these steps current.

Grok

On grok.com, connectors are added from the connectors page.

1. Go to grok.com/connectors, click New Connector and choose Custom.
2. Name it and enter the server URL. There is no key to paste.
3. Allow the sign-in pop-up, sign in, keep read, write and stay connected checked, and click Allow.
4. Ask Grok to list your workspaces to confirm.

A bot built on the xAI API uses remote MCP tools instead, with a server_url, a server_label and an authorization header. See the Grok integration page for both paths.

Cursor

Cursor reads MCP servers from ~/.cursor/mcp.json, or .cursor/mcp.json in a project, as its MCP documentation explains. Add the server by URL:

{
  "mcpServers": {
    "wavefront": { "url": "https://wavefront.adwave.com/mcp" }
  }
}

Cursor shows Needs login next to the server. Click it to finish signing in through the browser. The Cursor integration page covers team setups that use an API key from an environment variable.

How Wavefront works with agents

Wavefront is a remote MCP server, and the web app, the REST API and MCP share one core, so an agent sees the same workspaces, plans and drafts you do. Agents can talk to the strategist, edit the plan, commission and revise drafts, comment, read search data, and publish or schedule versions a person approved. They can’t approve, withdraw a page, post to social networks or send a newsletter. Content an agent writes or changes never qualifies for auto-approval.

The agents overview explains the model, and the help article on connecting an AI agent walks through setup. For the full technical reference, the MCP connection guide covers sign-in, scopes, API keys and troubleshooting, and the MCP tool reference lists every tool. Building something custom? Start with the developer overview.

Sources

1. Model Context Protocol: What is MCP?
2. Anthropic: Introducing the Model Context Protocol (November 25, 2024)
3. Model Context Protocol specification: Tools
4. Model Context Protocol specification: Authorization
5. OpenAI: Building MCP servers for ChatGPT and API integrations
6. Anthropic Help Center: Getting started with custom connectors using remote MCP
7. Cursor Docs: Model Context Protocol
8. xAI Docs: Remote MCP tools