Skip to content

Connect an AI agent (Claude Code, Codex)

Tiger is an MCP server: an AI coding agent such as Claude Code or OpenAI Codex can connect to your site and work in it — create pages, write blog posts, organize media, build menus — using the same permissions a signed-in user would have. This page gets you from zero to a connected agent.

Two kinds of work, and which tool does each:

You want to… Use
Change content and settings — pages, posts, menus, media, users MCP (this page). The agent calls your site's API; nothing touches your code.
Change code — build a custom module, a theme, a new feature Your agent working in your app's source code (a local checkout or a TigerDen server), following the repo's AGENTS.md.

1. Turn on the MCP endpoint

MCP is off by default — your site won't accept AI connections until you switch it on.

  1. Sign in as an admin and open Admin → Settings → MCP Server (/mcp/admin).
  2. Switch on Enable the MCP endpoint.

2. Create an access token

On the same screen, mint a token for the agent. Give it only what the job needs:

  • Modules — pick the areas the agent may touch (for example, CMS and Blog). The agent only sees tools from those modules.
  • Read-only — tick it if the agent should only look, never change anything. A write attempt is refused.
  • Org-scoped — the token acts for your organization rather than for you personally, so it keeps working if you leave and its actions are recorded against the token.

The token (tgr_…) is shown once. Copy it somewhere safe — treat it like a password.

3. Connect your agent

Tiger ships a small connector, mcp-bridge.php, that any MCP client can launch. It needs only PHP on your computer — no Node.js. Download it from the MCP Server screen (the Download mcp-bridge.php link) and save it somewhere permanent, for example ~/tiger/mcp-bridge.php.

Claude Code

Run once, in a terminal:

claude mcp add tiger \
  -e TIGER_MCP_URL=https://your-site.com/mcp \
  -e TIGER_MCP_TOKEN=tgr_your_token \
  -- php ~/tiger/mcp-bridge.php

Then start Claude Code and run /mcp — tiger should show as connected.

OpenAI Codex

Add this to ~/.codex/config.toml:

[mcp_servers.tiger]
command = "php"
args = ["/Users/you/tiger/mcp-bridge.php"]

[mcp_servers.tiger.env]
TIGER_MCP_URL = "https://your-site.com/mcp"
TIGER_MCP_TOKEN = "tgr_your_token"

Restart Codex; the Tiger tools appear in its tool list.

Other clients

Claude Desktop, Cursor and other MCP clients use the same three values: the command (php), the bridge path, and the two environment variables. The MCP Server screen shows a ready-to-paste configuration with your site's URL filled in.

4. Check it works

Ask the agent: "List the tools you have for Tiger." You'll see names like cms__page__save or blog__post__datatable — that's module__service__method, one tool per action the token allows.

Then try something small and visible: "Create a draft page called 'Hello from my agent'." Open Admin → Content and the draft is there.

A first project

A good first session, start to finish:

  1. "Create a page called About with a short bio, and publish it."
  2. "Upload this image to the media library and add it to the About page."
  3. "Add About to the main menu."
  4. "Write a blog post introducing the site, and add a comments thread to it." (Comments must be switched on first — see Comments & reviews.)
  5. "Change the About page's heading." — then check the menu, the blog and the mobile layout still look right. Every page save keeps a version, so you can restore an earlier one from the page editor's version history.

Keeping it safe

  • Least privilege. One token per job, with only the modules it needs. Prefer read-only when the agent is just researching.
  • Revoke freely. The MCP Server screen lists your tokens; revoking one takes effect immediately. To rotate, mint a new token, update your agent's configuration, then revoke the old one.
  • Staging first. Point a new or unfamiliar workflow at a staging copy of your site before your live one.
  • Everything is logged. Every tool call is recorded with the token that made it.
  • The agent can't exceed the token. Each call goes through the same permission checks as a person clicking in the admin; a token can never do more than its role and scope allow.

For how the server works under the hood, see MCP — how it works.