FlavourSuite AI documentation

Everything you need to connect AI agents β€” Claude, Cursor, ChatGPT, and any MCP client β€” to your WordPress site, safely.

Quick start

  1. Install and activate the plugin (WordPress 6.9+, PHP 7.4+).
  2. Go to Settings β†’ FlavourSuite AI and switch on β€œExpose this site to AI agents over MCP”. Until you do, the endpoint doesn't exist β€” installing changes nothing.
  3. Create an Application Password: Users β†’ Profile β†’ Application Passwords, give it a name like my-agent, and copy the generated password. Agents authenticate as this user and can only see and do what this user can.
  4. On the settings page, paste the Application Password into the snippet builder β€” it produces a finished configuration to copy into your MCP client (per-client guides below).
  5. Ask your agent something: β€œGive me an overview of my site.”
FlavourSuite AI settings page: master switch, per-tool toggles, and connection snippet

Every tool has its own switch. Read-only tools are on by default; anything that writes data (the propose tools) stays off until you enable it.

Connect your agent

Your endpoint is https://example.com/wp-json/flavoursuite-ai/mcp (shown on the settings page). Authentication is HTTP Basic with your WordPress username and the Application Password, sent as a header:

Authorization: Basic <token>
You don't need to build the token yourself. The settings page (Settings β†’ FlavourSuite AI β†’ Connect an agent) has a generator: paste your Application Password, click Build connection snippet, and copy the finished configuration β€” the token is computed in your browser and never sent anywhere. Command-line users can equivalently run echo -n 'yourname:xxxx xxxx xxxx xxxx' | base64 (the spaces WordPress shows in the password are fine β€” it ignores them).

Which client? If you're not a developer, use Claude Desktop β€” it's a one-time copy-paste into a config file, no terminal needed. Developers working in an editor will want Claude Code, Cursor, VS Code, or Codex, which all take the same URL + header.

Claude Desktop (best for non-developers)

  1. Install Node.js if you don't have it (the standard installer, all defaults).
  2. In Claude Desktop open Settings β†’ Developer β†’ Edit Config. That opens claude_desktop_config.json (on Windows: %APPDATA%\Claude\, on Mac: ~/Library/Application Support/Claude/).
  3. Paste this, with your site's URL and the token from the settings-page generator, save, and restart Claude Desktop:
{
  "mcpServers": {
    "flavoursuite": {
      "command": "npx",
      "args": [
        "mcp-remote",
        "https://example.com/wp-json/flavoursuite-ai/mcp",
        "--header",
        "Authorization: Basic <token>"
      ]
    }
  }
}

Your site's tools appear under the πŸ”Œ icon in new conversations.

Why the mcp-remote bridge? Claude's custom connectors UI (Settings β†’ Connectors β†’ Add custom connector) connects from Anthropic's cloud, not your machine, and authenticates with OAuth β€” so it can't send an Application Password header and can't reach local/intranet sites. The bridge above runs on your computer and works everywhere. A future FlavourSuite version may add OAuth so the connector UI works directly.

Claude Code

claude mcp add --transport http flavoursuite \
  https://example.com/wp-json/flavoursuite-ai/mcp \
  --header "Authorization: Basic <token>"

Verify with claude mcp list β€” the server should show βœ“ connected. Add --scope user to make it available in every project.

Cursor

Add to .cursor/mcp.json in your project (or ~/.cursor/mcp.json globally; also editable via Settings β†’ Tools & MCP):

{
  "mcpServers": {
    "flavoursuite": {
      "url": "https://example.com/wp-json/flavoursuite-ai/mcp",
      "headers": {
        "Authorization": "Basic <token>"
      }
    }
  }
}

VS Code (GitHub Copilot)

Copilot's agent mode supports MCP natively. Add to .vscode/mcp.json in your workspace (or your user profile's mcp.json) β€” note the top-level key is servers, not mcpServers:

{
  "servers": {
    "flavoursuite": {
      "type": "http",
      "url": "https://example.com/wp-json/flavoursuite-ai/mcp",
      "headers": {
        "Authorization": "Basic <token>"
      }
    }
  }
}

Codex (OpenAI)

Add to ~/.codex/config.toml (shared by the Codex CLI and IDE extension). Use http_headers, not bearer_token_env_var β€” WordPress expects Basic auth, not a Bearer token:

[mcp_servers.flavoursuite]
url = "https://example.com/wp-json/flavoursuite-ai/mcp"
http_headers = { "Authorization" = "Basic <token>" }

Verify by running /mcp inside a Codex session.

ChatGPT

ChatGPT adds remote MCP servers via Settings β†’ Connectors β†’ Advanced β†’ Developer Mode (Plus/Pro/Business plans), then Add custom connector with your endpoint URL. Two limitations to know: connections originate from OpenAI's cloud (your site must be publicly reachable over HTTPS), and its custom-connector auth supports OAuth or pasted tokens sent as Bearer β€” not the Basic scheme Application Passwords use. Until FlavourSuite ships OAuth support, use one of the clients above with ChatGPT-side workflows, or front your endpoint with an OAuth proxy.

Other clients (Cline, Windsurf, LibreChat, …)

Any MCP client that supports Streamable HTTP works: give it the endpoint URL and the Authorization: Basic <token> header β€” the JSON shape is usually identical to the Cursor example above.

What agents can do

ToolWhat it doesDefault
site-overviewWordPress/PHP versions, active plugins, content counts (administrators only)on
list-recent-postsRecent published entries of any public post typeon
search-contentFull-text search across public content, with snippetson
site-policiesPrivacy, terms, returns, shipping, FAQ and contact pages with their texton
get-custom-cssRead the current Additional CSSon
list-change-requestsStatus of the agent's proposals (pending / applied / rejected / rolled back)on
propose-cssPropose a change to Additional CSS β€” creates a pending change requestoff
propose-content-editPropose a new title/content for a post or page β€” same review flowoff
woo-store-overviewWooCommerce product/order counts, currency (appears when WooCommerce is active)on
fluent-crm/*CRM context, campaigns, automations (appears with FluentCRM 3.0+; no contact-level data)on

First prompts to try

Approvals: agent writes, your call

No tool writes to your live site. When an agent uses a propose tool, it creates a change request β€” a stored before/after pair β€” and gets back a review link. Nothing changes until a human approves it in wp-admin.

  1. Enable propose-css and/or propose-content-edit under Settings β†’ FlavourSuite AI.
  2. The agent proposes; the proposal appears under Tools β†’ Agent Changes as a red/green diff.
  3. You Approve & apply, or Reject. Applied changes keep a Roll back button that restores the exact before-state.
Tools β†’ Agent Changes: a pending CSS proposal shown as a before/after diff with Approve & apply and Reject buttons
Staleness protection. Each proposal records the state it was based on. If a human edits the CSS or the post after the proposal was made, applying is refused instead of overwriting that work β€” same for rolling back something that was edited after it was applied. When that happens, reject the proposal and ask the agent to propose again from the current version.

Content edits are additionally covered by WordPress revisions, so there's always a second way back.

Security model

Troubleshooting

The agent can't connect (401 / unauthorized)

The endpoint returns 404

A tool is missing from the agent's list

Applying a proposal says it's stale

That's the safety net working: the CSS or post changed after the agent proposed. Reject the request and ask the agent to propose again from the current state (it will re-read via get-custom-css or the content tools).

Something not covered here? Open an issue on GitHub.