ClaudeMap

Yandex Metrika MCP

A read-only MCP server for Yandex Metrika (Яндекс Метрика): ask about traffic, sources, landing pages and goals, and every answer carries dates, the counter's time zone, sampling status and provenance — browser sign-in, tokens kept out of the chat in the OS credential store, strict accuracy that refuses sampled data, 13 tools and 3 prompts, full CSV/JSON exports in the background (MIT)

MCP Serversmcp serveryandex metrikaanalyticsread onlyseodata export

DevCrabs' Yandex Metrika MCP lets Claude, Cursor, VS Code, Codex or any MCP client query Yandex Metrika reports directly (129 stars as of writing, 2026-11-12): v1.0.0 is the first stable release, with tool names, arguments and outputs under semantic versioning (npm latest is 1.0.0; Node.js 22.13 or later). Sign-in goes through the browser — no token pasting, no OAuth app of your own; the token lives in the OS credential store (macOS Keychain, Windows Credential Manager, Linux Secret Service) and never enters the chat; without a credential store the server refuses to store tokens unless you explicitly choose an unencrypted user-only file. The numbers are built to be checkable: calendar periods in the counter's time zone, complete days by default, strict accuracy that refuses sampled data instead of calling it exact, API totals instead of sums of visible rows — every answer carries exact dates, the time zone, sampling status, filters and provenance. 13 read-only tools and 3 prompts (weekly_review, traffic_drop, landing_conversion); no write endpoints, no raw HTTP tool; multiple accounts connect in isolation with nothing shared between them; full CSV/JSON exports run as background jobs with checksums, a manifest and resume after a restart. Privacy defaults strip URL query strings from labels and send no telemetry; the report data you request goes to your AI client and its model provider — the token never does.

What it does

  • Verifiable numbers: calendar periods in the counter's time zone, complete days by default, strict accuracy that refuses sampled data instead of calling it exact, API totals instead of sums of visible rows
  • Browser sign-in: no token pasting, no OAuth app of your own; tokens live in the OS credential store (macOS Keychain, Windows Credential Manager, Linux Secret Service) and never enter the chat or the model provider
  • Small, predictable, read-only surface: 13 tools and 3 prompts (weekly_review, traffic_drop, landing_conversion); no write endpoints, no raw HTTP tool; tool names, arguments and outputs follow semver since v1.0.0
  • Several accounts and clients: isolated connections with their own counters, nothing shared between them
  • Full exports: background CSV/JSON export of every row, with checksums, a manifest and resume after a restart
  • Privacy defaults: URL query strings removed from labels, no telemetry; the synthetic conn_demo connection lets you try reports without an account

When to use it

For anyone checking Yandex Metrika from Claude Code, Cursor or VS Code: comparing organic landing pages across two weeks, weekly reviews, investigating a traffic drop (the traffic_drop prompt), exporting a month of traffic sources to CSV for Excel; agencies can keep each client's counter on its own isolated connection.

Getting started

# requires Node.js 22.13 or later; no API key and no OAuth app of your own

# one command (writes the client config and keeps a backup):
npx -y @dev-crabs/yandex-metrika-mcp setup --client claude-desktop   # or claude-code, cursor, vscode, windsurf, codex

# Claude Code:
claude mcp add --scope user yandex-metrika -- npx -y @dev-crabs/yandex-metrika-mcp

# any MCP client (VS Code uses "servers" with "type": "stdio"):
{ "mcpServers": { "yandex-metrika": { "command": "npx", "args": ["-y", "@dev-crabs/yandex-metrika-mcp"] } } }

# Windows: some clients cannot start npx directly — setup writes "command": "cmd", "args": ["/c", "npx", "-y", "@dev-crabs/yandex-metrika-mcp"] there
# Claude Desktop can also install the .mcpb bundle from the GitHub release

# restart the client and ask: "Connect Yandex Metrika and compare the last two weeks by traffic source"
# the assistant calls metrika_connect and gives you a local sign-in link; the synthetic conn_demo connection works without an account

Things to know

Platform status is stated in the README, not assumed: the server is covered by tests against synthetic data and a simulated API, and the live Yandex sign-in and report checks were done on macOS — other clients and platforms are tracked in docs/release-checklist.md, so treat them as unverified until you check your own. Tokens live in the OS credential store; without one, the server refuses to store tokens unless you explicitly choose an unencrypted user-only file. The privacy boundary runs both ways: the report data you request goes to your AI client and its model provider — tokens never do. The repo carries a Russian documentation set (README.ru.md and docs/ru) beside the English README. MIT © DevCrabs. 129 stars as of writing (2026-11-12).

Repository

DevCrabs-Inc/yandex-metrika-mcp

Charted

Related on the map

  • Blitz Strike

    MCP Server

    A penetration-testing methodology delivered as one MCP server: BLITZ maps the attack surface, EAGLE-EYE traces source-to-sink, STRIKE live-verifies every finding before it is reported.

    • mcp-server
    • security
    • pentest
    • methodology
  • Ripwire

    MCP Server

    Red Hat's "ripgrep for AI context" — a C++23 zero-dependency CLI that also speaks MCP.

    • mcp
    • code-context
    • cli
    • red-hat
  • Deja Vu

    MCP Server

    One shared session memory for 25+ coding agents — Claude Code, Codex, and Cursor finally remember the same project.

    • mcp
    • memory
    • cross-agent
    • context