MCP: connect AI assistants

The control center is an MCP server (Model Context Protocol). Claude Code, Claude.ai, ChatGPT or your own agents use it to read status, maintain categories and copy, edit templates, deploy and roll back.

Endpoint: https://affiliateai.evnxt.de/mcp (Streamable HTTP, JSON responses, stateless).

Access #

• Personal access token: create it under MCP und Tokens with name, expiry (1 to 365 days), sites and permission levels. It is shown once and stored only as a hash.
• OAuth 2.1: Claude.ai, ChatGPT and Claude Code can sign in without a token (dynamic client registration, PKCE required, consent page for sites, permissions and lifetime). Revoke access at any time.

Connect #

Claude Code with a token:

claude mcp add --transport http affiliateai https://affiliateai.evnxt.de/mcp \
  --header "Authorization: Bearer YOUR_TOKEN"

With OAuth, omit --header, then run /mcp in Claude Code and choose Authenticate.

Claude.ai and Claude Desktop: Settings, Connectors, add custom connector, URL as above.

ChatGPT: Settings, Apps and Connectors, Advanced, enable developer mode, create connector, URL as above, authentication OAuth.

OpenAI Responses API:

{
  "type": "mcp",
  "server_label": "affiliateai",
  "server_url": "https://affiliateai.evnxt.de/mcp",
  "headers": {"Authorization": "Bearer YOUR_TOKEN"},
  "require_approval": "always"
}

Permission levels #

Effective rights are the intersection of token levels, your account role (members may only read) and the sites granted.

Tools (42) #

Overview: sites_auflisten, site_status. Content: 19 tools for profile, categories, brands, filters, page copy, search jobs and exclusions. Data: produkte_suchen, produkt_lesen, preisalter, statistik, job_starten, job_status. Code: 10 tools for listing, reading, writing, patching and reverting overlay files. Deploy: vorschau_pruefen, deploy_starten, deploy_status, deploy_verlauf, zurueckrollen. The full list is in the German version.

Security model #

• Code tools only touch your overlay, never the core, no dot files, symlinks or .., UTF-8 text up to 200 KB.
• Every change is a Git commit marked “via MCP” and can be reverted.
• Every MCP deploy passes a preview check first. See deploy.
• Rate limit of 120 calls per minute per token.
• Every tool call is written to the audit log, content only as length and checksum.