Yandex Metrica Mcp

yandex-metrica-mcp

[CI] [npm] [License: MIT]

Ask your Yandex Metrica analytics in plain language — from Claude, Cursor, or any MCP client.

A Model Context Protocol (MCP) server for Yandex Metrica. It lets an AI agent query your web-analytics data — traffic, sources, landing pages, conversions, geography, devices and trends — through a small set of flexible, read-only tools — plus opt-in write tools that you confirm one change at a time.

Read-only by default, no secrets stored in the package: interactive login uses a built-in public OAuth client with PKCE, and the server talks only to Yandex.

General-purpose: marketing, product analytics, e-commerce and SEO are all fair game — the server is not tied to any one use case.

Demo

Point an AI agent at your counter and ask about your traffic — the server queries Yandex Metrica and hands back real, read-only data, no dashboards. Here the run_report tool answers a “traffic sources, last 7 days” question against a live counter:

[yandex-metrica-mcp querying live Yandex Metrica traffic sources, read-only]

Quickstart

1. Add the server to your MCP client (e.g. Claude Desktop) — no token required up front; you log in interactively in step 2:

{    "mcpServers": {        "yandex-metrica": {            "command": "npx",            "args": ["-y", "yandex-metrica-mcp"],            "env": { "YANDEX_METRIKA_COUNTER_ID": "12345678" }        }    }}

2. Log in once — one command, no app registration, no secret stored:

npx yandex-metrica-mcp auth

Approve access in the browser and you're done — the code is handed back automatically over a local redirect, no copy-paste. The login uses authorization-code + PKCE, so no client secret ever touches your machine; the token is cached (mode 0600) and valid for ~1 year.

3. Ask your agent about traffic, sources, conversions, geography, devices, or trends — see Examples for prompts.

Prefer a static token (CI / non-interactive) or your own OAuth app? See Authentication.

Or install as a Claude Code plugin

The repo doubles as the plugin marketplace for BoxLab's Yandex MCP family, so you can install the server through Claude Code's plugin system instead of the config above:

/plugin marketplace add BoxLab-Ltd/yandex-metrica-mcp/plugin install yandex-metrica-mcp@boxlab

Then sign in from the chat with the login tool (or run npx yandex-metrica-mcp auth). The same marketplace also carries yandex-webmaster-mcp. To turn on the write tools in the plugin, start Claude Code with YANDEX_METRIKA_ENABLE_WRITE=1 in the environment.

Or install as a Claude Desktop extension (.mcpb)

For a one-click install with no JSON, download the .mcpb from the latest release and open it with Claude Desktop (or drag it into Settings → Extensions). It asks for an optional default counter id; then sign in from the chat with the login tool (or run npx yandex-metrica-mcp auth).

Why

There is no official Yandex Metrica MCP server, and existing community ones are mostly thin, unmaintained, or dump raw data straight into the model's context. This server aims to be the well-engineered, well-maintained, open option: flexible report tools, strict token/context discipline, read-only by default.

Features

• run_report — flexible wrapper over the Reporting API (/stat/v1/data).
• run_comparison — compare two periods with absolute and percentage deltas.
• run_drilldown — drill down through a dimension tree.
• run_timeseries — metrics split into a time series (/bytime) for trends.
• get_metadata — discover the counters on your account and a catalog of common dimensions/metrics (and Logs API fields) so the model queries with real names.
• describe_counter — read one counter's configuration (goals, segments, filters, operations, access grants, chart annotations) via an include selector. The goals section gives the goal ids needed for conversion metrics in run_report.
• logs_request / logs_status / logs_download / logs_clean — Logs API: export raw, un-sampled session (visits) or hit (hits) rows. Async lifecycle (request → poll → download → clean); logs_download returns a bounded sample inline by default, or streams the full export to a file — never dumping raw rows into the model's context.
• login / submit_code — sign in to Yandex Metrica from your MCP client, no terminal needed: login opens the browser and captures the code over a local redirect, or hands back a URL and takes the pasted code via submit_code.
• Built-in context control: field selection on by default, low default row limits, and sampling/quota surfaced back to the model.
• Analysis hygiene: the server hands every client a short set of rules (exclude robots, compare complete periods, never sum users, use explicit goals, …). The Claude Code plugin adds the fuller metrica-hygiene skill.

Write tools (opt-in)

Off unless you set YANDEX_METRIKA_ENABLE_WRITE=1. Then the server adds:

• manage_annotation — create, update or delete a note pinned to a chart date (a release, a campaign start, an outage).
• upload_expenses — import ad costs from a non-Yandex source as CSV, so Metrica reports cost and ROI. dryRun: true validates the file and shows the rows, dates and totals without sending anything.

With the flag on, sign-in also asks for the metrika:write scope — run auth (or the login tool) once more after turning it on. Without the flag the sign-in asks for read access only, and tokens issued earlier keep working. Every change is confirmed by you before it is sent: your client shows a confirmation prompt if it supports MCP elicitation, otherwise the tool returns a summary and a one-time token that the agent can use only after you agree. The server also checks that your account has edit access to the counter.

Local HTTP mode

By default the server talks stdio, started by your MCP client. For clients that connect by URL, or to share one running server between several clients, start it over Streamable HTTP:

npx yandex-metrica-mcp --http            # http://127.0.0.1:7801/mcpnpx yandex-metrica-mcp --http --port 9000

Or set MCP_TRANSPORT=http (and MCP_HTTP_PORT). Every client gets its own session; they share one sign-in and one request budget. The server listens on 127.0.0.1 only and rejects requests whose Host or Origin is not local, so a web page cannot reach it. To listen on another interface (--host 0.0.0.0, e.g. in Docker) you must set MCP_HTTP_TOKEN; clients then send Authorization: Bearer <token>. This is a local mode, not hosting: one user, one Yandex account.

Requirements

• Node.js >= 18
• Yandex Metrica credentials with the metrika:read scope (see Authentication), plus metrika:write for the opt-in write tools. Whoever the credentials belong to must have access to the counters you query.

Authentication

Recommended: interactive login. No app registration needed — the server ships a built-in public OAuth client. Run once:

yandex-metrica-mcp auth     # or, in dev: bun run auth

It opens the Yandex consent page in your browser; after you approve, the code is returned automatically over a loopback redirect (http://127.0.0.1:53682) — no copy-paste. If that port is taken, it falls back to showing a code you paste in (force that flow with auth --oob, or change the port with YANDEX_OAUTH_LOOPBACK_PORT). The token is cached at ~/.config/yandex-metrica-mcp/token.json (mode 0600) and is valid for ~1 year; re-run auth when it expires. The login uses authorization-code + PKCE, so no client secret is stored anywhere. A cached login takes precedence over YANDEX_METRIKA_TOKEN.

From your MCP client (no terminal). Not signed in yet? The server still starts — ask your agent to run the login tool and it does the same browser flow in-process (or returns a URL and takes the code via submit_code). Handy for GUI clients like Claude Desktop, where there is no terminal to run auth.

Static token. Get a token for an app with the metrika:read scope at https://oauth.yandex.ru and pass it as YANDEX_METRIKA_TOKEN — handy for CI or non-interactive use.

Own OAuth app (optional). To use your own app instead of the built-in one, set YANDEX_OAUTH_CLIENT_ID; add YANDEX_OAUTH_CLIENT_SECRET to also enable automatic token refresh.

Configuration

The Quickstart covers the happy path. For all options — static token, your own OAuth app, default counter, request tuning, language — see .env.example. The published package runs on Node (so npx/MCP clients work out of the box); local development uses Bun.

Examples

Once connected, an agent can answer questions like:

• “How many visits and users did counter 12345678 get last week, split by traffic source?” → run_report with metrics: ["ym:s:visits","ym:s:users"], dimensions: ["ym:s:lastsignTrafficSource"].
• “Compare this week's organic conversions to last week's.” → run_comparison (server returns A, B, and the deltas).
• “Which operating systems do my visitors use? Let me drill into Windows versions.” → run_drilldown, then again with parentId.
• “What counters do I have?” → get_metadata.
• “List this counter's goals, then show how the ‘Purchase’ goal converted last week.” → describe_counter (include: ["goals"]) for the goal id, then run_report with ym:s:goal<id>conversionRate.
• “Export last month's raw sessions with landing pages and referrers for offline analysis.” → logs_request (source: "visits"), poll logs_status, then logs_download (mode: "file"), then logs_clean to free the quota.

Development

This project is Bun-first:

bun installbun run dev        # run from source with hot reloadbun run typecheck  # tsc --noEmitbun run lint       # eslintbun test           # bun's test runnerbun run build      # emit dist/ with tsc (Node-compatible)

The skill has an eval suite in evals/ that runs against mocked Metrica tools, so it never touches the real API:

claude plugin eval . --ablation none

License

MIT © boxlab