# Yandex Metrica MCP > 13 MCP tools over stdio for Claude Code, Codex and Cursor: Metrica reports, raw Logs API export, counters and goals. Run with Docker, MIT. Source code: https://github.com/stufently/yandex-metrica-mcp License: MIT - 13 MCP tools covering counter management, goals, the Reporting API and the full Logs API lifecycle. - Reporting API queries with built-in period comparison: pass date1_b/date2_b to get metrics for segments A and B. - Logs API raw export: evaluate, create, poll, download parts and clean, with a 50KB preview by default. - Eight tools only read; five change the account, and each tool declares readOnlyHint and destructiveHint. - Runs locally over stdio from a Docker image on GHCR — no web server, no persistent storage, no daemons. - Token via environment variable only, never logged; the image runs as a non-root user. ## Documentation - [yandex-metrica-mcp](https://stufently.github.io/yandex-metrica-mcp/docs/): MCP server for the Yandex Metrica Reporting API, Logs API and counter management, designed for use with Claude Code and Codex CLI. Eight of its 13 tools… ### How do I install the Yandex Metrica MCP server? Use Docker: docker run -i --rm -e YANDEX_METRICA_TOKEN ghcr.io/stufently/yandex-metrica-mcp:latest. Nothing to clone and no Node.js needed. The npm package @stufently/yandex-metrica-mcp is not published yet. ### Which token do I need? A Yandex OAuth token with the metrika:read scope in YANDEX_METRICA_TOKEN. Creating and deleting counters requires the metrika:write scope. ### Which AI clients does it work with? Any MCP client over stdio. The README has configs for Claude Code, Claude Desktop, Cursor, Windsurf, Codex CLI and Zed; Claude Code on Ubuntu is verified in daily use, the others are documented only. ### Is the server read-only? No. Eight of the 13 tools only read. Five change your Metrica account: create and delete a counter, and create, cancel and clean a Logs API export. ### How does the Logs API export work? It is asynchronous: evaluate, create a request, poll yandex_metrica_get_logs_request until status is processed, download each part, then clean. Parts return a 50KB preview unless you set mode to full. ### Does it cache responses? No. Every tool call hits the Yandex API. Retries with exponential backoff are built in: 3 retries on statuses 429, 500, 502, 503 and 504. ## Optional - [Full text of all docs](https://stufently.github.io/yandex-metrica-mcp/llms-full.txt)