Skip to main content

MCP — Connect Your AI

RIFT speaks MCP (Model Context Protocol), the standard that lets AI assistants — Claude, Cursor, and any MCP-capable agent — work with your tools. Connect once, and your AI can explore your content, edit pages, and propose changes — all inside RIFT's guardrails.

The model

Once connected, you can ask your AI things like:

"Update the homepage: add two new feature cards for the programs below, and create a page for each one. Keep the design system. Then send it for review."

Your AI does the work inside RIFT's guardrails: everything it touches lands in a change-set, gets built to your staging site, and waits for a person on your team to review and approve it in RIFT before anything goes live. An AI connected to RIFT cannot publish, archive, or administer — ever. Every action it takes is recorded in the same audit trail as a human edit.

Create an agent key

In RIFT: API Keys → Create key, choose role agent, and copy the key (it is shown once). The agent role is the least-privilege grant built for AI callers — it can read and propose, and nothing else. Revoke it any time from the same page.

Your MCP endpoint is your RIFT URL plus /api/mcp:

https://<your-rift-host>/api/mcp

Connect your client

Claude Code

claude mcp add --transport http rift "https://<your-rift-host>/api/mcp" \
  --header "Authorization: Bearer <your-agent-key>"

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "rift": {
      "url": "https://<your-rift-host>/api/mcp",
      "headers": { "Authorization": "Bearer <your-agent-key>" }
    }
  }
}

Anything else

RIFT's MCP server is a standard remote server (Streamable HTTP). Any client that can send an Authorization: Bearer header works; point it at the URL above.

To verify: ask your AI "List my RIFT sites." If it answers with your site, you're connected.

Things to ask for

Explore your content

"What pages on our site talk about volunteer opportunities?"
"Show me the current homepage content."
"What components does our design system have?"

Update a page

"On the About Us page, rewrite the second section to mention our new office. Keep the existing structure and styles, then propose it for review."

Live pages are safe to ask about: RIFT automatically gives the AI a staged copy to edit — the live page doesn't change until your team approves.

Add new pages

"Create a page under Get Involved for our new mentorship program — here's the content: … Link to it from the Get Involved landing page, and use the same card component the other programs use."

Check on work

"What's in the change-set you just proposed? Give me the staging link again."

Or hand RIFT the whole job (the one-prompt path)

For self-contained requests, your AI can pass the entire job to RIFT's own content agent with the request_content_change tool — RIFT plans the work, writes the pages in your site's design system, and proposes the change-set itself, while your AI just polls until the staging and review links are ready:

"Ask RIFT to handle this end-to-end: add a 'Mentorship Program' page under Get Involved built from the following content: …"

The internal agent runs on your RIFT instance's own AI configuration and budget, under the same rules as any connected AI: staging only, everything in a reviewed change-set.

What happens when the AI finishes

When the work is done, the AI proposes the change-set and hands you two links:

  • Staging URL — see every change rendered on your real staging site.
  • Review URL — the change-set in RIFT: the original request, a before/after diff of every page, a link-resolution check on every reference, and one decision — approve everything, approve with specific pages excluded, or reject (which returns the work to draft). A comment is required either way, and the whole decision is recorded under the reviewer's name.

Approving can also publish the site in the same step (a checkbox, on by default). Until someone approves, your live site is untouched. That's the RIFT model in one line: agents propose, humans approve.

Good to know

  • One request, one change-set. Ask the AI to keep related edits in a single change-set so your reviewer sees them as one unit.
  • Links stay honest. The AI is told, before it links to anything, whether that link will actually resolve on your published site — RIFT's link graph does the checking.
  • Everything is attributable. Each agent action appears in the audit log under its API key and change-set. Revoking the key ends its access instantly.

RIFT also ships two playbook prompts your MCP client can surface directly (update-content and one-prompt-change) — pre-written instructions that walk any AI through the right sequence for a content update.

The MCP server is built on the same versioned REST API that powers everything else in RIFT — same access model, same audit trail. See the API Reference.