@advertising-chat/mcp-manage

Operate your Advertising.chat account from Claude Desktop, Claude Code, or any MCP-capable agent host — with your own account and permissions.

mcp-manage is a local MCP (Model Context Protocol) server that projects the Advertising.chat management API as ~39 tools: ad units, line items, creatives (including flow creatives), campaigns, ad groups, API keys, and real-time performance queries. Every call runs with your token and your role — the server can't do anything you couldn't do with curl.

Quick start

1. Authenticate once with the Advertising.chat CLI:

   npx ads-cli login   # or: ads login, if you have ads-cli installed

   mcp-manage picks up (and auto-refreshes) the cached token at ~/.ads-cli. Alternatively set ADCHAT_TOKEN to an access token.

2. Add the server to your MCP host. Claude Desktop (claude_desktop_config.json):

   {
     "mcpServers": {
       "advertising-chat-manage": {
         "command": "npx",
         "args": ["-y", "@advertising-chat/mcp-manage"]
       }
     }
   }

   Claude Code:

   claude mcp add advertising-chat-manage -- npx -y @advertising-chat/mcp-manage

3. Ask your agent things like "how many impressions did we serve yesterday?" or "create a disabled line item for the homepage ad unit with the summer-sale creative."

Configuration

| Env var | Meaning | |---|---| | ADCHAT_TOKEN | Explicit access token (overrides the ~/.ads-cli cache). | | ADCHAT_ORG | Target org id. Default: your own org, resolved from your identity. | | ADCHAT_SYSTEM_REQUEST | true to send X-Adchat-System-Request (sysops only; requires ADCHAT_ORG). | | ADCHAT_MANAGE_BASE_URL | Management API base URL override (default https://api.advertising.chat/system). |

Safety model

• Your authority, nothing more. Requests carry your Auth0 token and target org; org membership and role checks happen server-side exactly as they do for the console. A viewer role can read everything and mutate nothing.
• Created disabled. Ad units, line items, campaigns, and ad groups are created with enabled: false unless you explicitly ask to launch.
• Deletes need confirmation. Destructive tools refuse to run without an explicit confirm: true argument, and MCP hosts additionally prompt on non-read-only tools.
• Attributed. Every request is tagged X-Adchat-Agent: mcp-manage for audit.
• The server never logs your token. Installing third-party MCP servers alongside this one is your host's trust decision.

Embedding the tool layer

The tool definitions and handlers are importable without the stdio server — the Advertising.chat console copilot consumes the same table server-side:

import {
  tools,            // definitions: name, description, zod schema, annotations
  createHandlers,   // bind handlers to an HTTP client
  createHttpClient,
  validateFlowPayload,
} from '@advertising-chat/mcp-manage';

const http = createHttpClient({ getToken, org });
const handlers = createHandlers(http);
await handlers.list_ad_units({ limit: 10 });

Docs

Full tool catalog and walkthrough: https://docs.advertising.chat/adkit/manage-mcp/