Infostart connects Direct, Metrica and Wordstat to ChatGPT: 6 commands, shared access
Infostart has connected Direct, Metrica and Wordstat to ChatGPT. A marketer types "show active campaigns with conversions" in the chat and gets a table without opening any of the dashboards. The link runs through an MCP server hosted on ChatGPT Sites, with six read-only commands. The catch is shared tokens: the server does not check an employee's personal Yandex permissions, so departments with separate ad accounts need separate Sites.
AI-processed from Habr AI; edited by Hamidun News
Previously, to match ad spend against conversions, a marketer had to open the Yandex Direct dashboard, then Metrica, then Wordstat, and stitch the numbers together by hand. The Infostart team removed that switching. Now it is enough to type "analyze the active ad campaigns and pull in conversions from Metrica" in ChatGPT and get a table with spend, clicks and goal completions.
To make this work, the team built an MCP server (MCP is a protocol that lets an AI chat request data from external services) and hosted it on ChatGPT Sites, that is, on hosting inside ChatGPT itself. Direct is Yandex's ad campaign platform, Metrica is web analytics, and Wordstat is search query statistics. This is worth reading if you are thinking about connecting work data to an AI chat: the setup is convenient, but it has one notable catch.
The author of the Habr article estimates that they spend about 70% of the workday in AI tools, and with every new MCP connection there are fewer reasons to open other interfaces.
What the plugin can do
An employee installs the plugin from the GPT catalog: finds it, clicks "Install" and "Connect," then selects it in the chat. No local programs or personal Yandex keys are needed. The server pulls data from the Yandex API, and ChatGPT only explains the result.
There are six tools:
- direct_list_campaigns: campaigns, with their statuses, types and dates;
- direct_report: impressions, clicks, spend and conversions for a selected period;
- metrika_goals: the list of goals for a selected counter;
- metrika_report: aggregated traffic metrics and goal completions;
- wordstat_top: popular and related queries with frequency;
- wordstat_dynamics: historical dynamics of query frequency.
The authors warn that the model's conclusions and the selected goals need to be checked. A goal completion in Metrica does not, by itself, confirm a sale.
How the server was built and published
The server is written in the company's repository. The code fixes the available commands and validates parameters, accounts and counters, so it is not possible to pass an arbitrary method or API address. The Direct and Metrica tokens and the Wordstat key are stored in the Site's server variables. They do not end up in the chat, in the source code, or in tool arguments. The server accepts requests at /mcp, and when publishing, the owner gets a linked plugin. A separate OAuth server for signing in from ChatGPT was not needed.
The team checked the connection on 17 November 2026 via check_connection: reading from Direct and Metrica succeeded. For Wordstat, they tested only the free regions reference, so this test does not confirm that the frequency data is fresh. When moving a ready-made MCP server to Sites, you need to account for the limitations of their environment, and paid external APIs are billed separately.
Where the catch is
The Site owner chooses the audience: invited users, the workspace, or access by link. Infostart chose its own members, and the plugin is distributed through the internal catalog. View permission is enough for a marketer; no additional editors were added. The plugin has the Use my default setting and inherits the general Allow low-risk actions rule, which determines when ChatGPT asks for confirmation.
The main limitation is data. Access to the Site gives an employee the entire shared dataset of the integration. The server calls Yandex with shared tokens and does not check a person's personal permissions in Yandex. Suppose a company has two departments with different ad accounts: both will see everything. The fix is separate Sites or access separation in code.
This MCP has no commands for changing ads. If they become necessary, the authors advise building a separate MCP server and opening it only to those who are allowed to apply edits. Yandex token permissions are configured separately from ChatGPT permissions.
There is also dependence on a single vendor: the model, the working interface and the MCP hosting belong to one company. The classic dashboard will still be needed if the numbers do not match, the integration fails, or a needed function is missing. The authors sum up: when a company's data and tools are accessible from ChatGPT, it starts to serve as a work OS.
- Nov 19, 2026Amazon and Google's Anthropic Stakes Valued at $308 Billion in IPO at $83 a Share
- Nov 19, 2026Google's Gemini 4 Neon runs twice as fast as Carbon and costs 17% less per task
- Nov 19, 2026OpenAI's coding agents would cost $2400 a day per researcher at API prices
- Nov 19, 2026Smysl.AI: An Agentic RAG Over Company Documents, Built in Kotlin Without Python
Key Questions Answered
What the plugin can do?
An employee installs the plugin from the GPT catalog: finds it, clicks "Install" and "Connect," then selects it in the chat. No local programs or personal Yandex keys are needed. The server pulls data from the Yandex API, and ChatGPT only explains the result.
How the server was built and published?
The server is written in the company's repository. The code fixes the available commands and validates parameters, accounts and counters, so it is not possible to pass an arbitrary method or API address. The Direct and Metrica tokens and the Wordstat key are stored in the Site's server variables. They do not end up in the chat, in the source code, or in tool arguments. The server accepts requests at /mcp, and when publishing, the owner gets a linked plugin.
Where the catch is?
The Site owner chooses the audience: invited users, the workspace, or access by link. Infostart chose its own members, and the plugin is distributed through the internal catalog. View permission is enough for a marketer; no additional editors were added. The plugin has the Use my default setting and inherits the general Allow low-risk actions rule, which determines when ChatGPT asks for confirmation. The main limitation is data. Access to the Site gives an employee the entire shared dataset of the integration.
Build & Ship Production AI Agents Without Boilerplate Code
Start your free 72-hour trial at Hamidun Academy (academy.hamidun.com) — interactive lessons on Claude Code & Cursor, live Tech Radar, and production agent templates.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.