Ezstat Mcp Server

EzStat MCP Server

A standalone Model Context Protocol server for EzStat. Let Claude / Cursor / any MCP-compatible agent push and read your metrics natively — no copy-paste, no dashboard hop. This is the product behind the "metrics your agents write and read themselves" positioning: every agent that touches your code can also touch your observability.

The server speaks stdio (the standard transport for local agent runners) and — as a hosted endpoint at https://mcp.ezstat.dev — streamable HTTP, exposing the same seven focused tools. Each tool description is written for an agent audience so the model knows when to call it.

Hosted endpoint — zero install

Agents can connect to the hosted streamable-HTTP endpoint with nothing but an EzStat API key — no npm install, no local process:

{  "mcpServers": {    "ezstat": {      "type": "http",      "url": "https://mcp.ezstat.dev",      "headers": {        "Authorization": "Bearer ezkey_your_api_key_here"      }    }  }}

• Multi-tenant: every request carries your key as Authorization: Bearer <ezkey>; the server holds no key of its own and each connection only ever sees its own account.
• Discovery is free: initialize and tools/list work without a key, so clients and registry inspectors can introspect the tool surface before you configure anything.
• Tool calls require the key: a call without one (or with an invalid one) returns a clean in-band MCP tool error telling the agent what to send — never a hang or a crash.
• Stateless: no sessions are issued or required; each POST is self-contained.

This hosted endpoint also satisfies the requirements for hosted/remote listings on MCP registries (e.g. Smithery's hosted-server path) — see server.json's remotes entry.

The stdio transport below remains fully supported and is still the right choice when you want the server running locally under your own runner.

Tools

Tool descriptions are tuned for agent reasoning — see src/server.ts.

Install (Claude Desktop / Claude Code / Cursor)

One-line install (npm/pnpm)

# pnpmpnpm add -g ezstat-mcp-server
# npmnpm install -g ezstat-mcp-server

Configure the MCP client

Claude Desktop (~/.config/Claude/claude_desktop_config.json on macOS/Linux, %APPDATA%\Claude\claude_desktop_config.json on Windows):

{  "mcpServers": {    "ezstat": {      "command": "ezstat-mcp-server",      "env": {        "EZSTAT_API_KEY": "ezkey_your_api_key_here"      }    }  }}

Cursor (~/.cursor/mcp.json):

{  "mcpServers": {    "ezstat": {      "command": "ezstat-mcp-server",      "env": {        "EZSTAT_API_KEY": "ezkey_your_api_key_here"      }    }  }}

Get your API key from https://ezstat.dev (it's the "ezkey").

Smithery

npx -y @smithery/cli install ezstat-mcp-server --client claude

Or visit https://smithery.ai/server/ezstat-mcp-server and click Install.

About EzStat

EzStat is a dead-simple hosted metrics service: one HTTP POST in, a live chart out. No SDK, no collector daemon, no YAML, no dashboard builder. Counters and value/gauge stats auto-create on first POST; charts, share/embed, and CSV/JSON export are built in, and Ask-Your-Data (live) answers plain-English questions from your own metrics, with every number verified against the retrieved data before it reaches you.

Flat monthly pricing by tracked stats — $19 / $49 / $149 — plus a card-gated free tier (no charge; see your data live before paying). Your data is yours: export any time, cancel any time. Anomaly detection and metric correlations are live. Features that are not shipped yet (alerts, weekly digest) are marked coming soon on the site rather than sold — what you see live is what works.

Coming from StatHat?

If your api.stathat.com calls stopped and your dashboards went dark: EzStat speaks StatHat's wire format — the same /ez, /c, /v endpoints, same params, same response. The migration is usually one line:

# beforecurl -X POST https://api.stathat.com/ez -d "stat=messages sent" -d "ezkey=KEY" -d "count=1"# after — change the host, use your EzStat key; stats auto-createcurl -X POST https://api.ezstat.dev/ez -d "stat=messages sent" -d "ezkey=EZSTAT_KEY" -d "count=1"

One gotcha, and it is the important one: check the response BODY, not the status code. Because the wire format is StatHat's, a rejected write (stale key, wrong key, quota hit) still returns HTTP 200, with the verdict in the body — {"status":200,"msg":"ok"} means recorded, {"status":"error","msg":"..."} means it was not. An unmodified StatHat client reports those as successes while your charts keep rendering old data, so nothing looks broken. New code should send X-EzStat-Strict: 1 on /ez and get real HTTP status codes instead. (This server already does the right thing: it treats an error body as a failure regardless of the HTTP code.) After switching, confirm points are arriving with ezstat.dev/tools/ezstat-verify.py — documented at ezstat.dev/docs#verify.

Most StatHat client libraries take a base-URL override in one line. Saved a StatHat CSV/JSON export? The importer recreates your stats and backfills history (8 MB / 500k points per file). Full guide: docs/stathat-migration.md · ezstat.dev/migrate/stathat — including the migrator deal: free tier to see it live first, a 12-month price-lock, and white-glove import for the first 25 migrations.

Why agent-native metrics

Your coding agent deploys, tests, and ships — it should also be the one tracking and reading the numbers. That's this server: metrics your agents write and read themselves. The reasoning: docs/agent-native-metrics.md · comparison with StatHat/StatFlow/Datadog: ezstat.dev/vs.

Environment variables

A starter .env.example is shipped.

How it talks to EzStat

The tools map to real EzStat API routes (all auth by EzStat API key):

• track_metric calls the StatHat-compatible EZ ingest ({"ezkey","stat","count"?,"value"?,"t"?}). Counter semantics: omit both count and value to record count=1 (counter +1).
• ask_ezstat calls Ask-Your-Data ({"query":"..."} → {answer, data, intent, ...}). This is the agent read path — count it as an agent read for usage tracking.
• read_stat returns the stat's recent series + summary; pass from/to (Unix seconds, converted to ISO 8601 at the HTTP boundary) to bound the window. The bucket width is auto-derived from the requested span (≤6h → minute, ≤7d → hour, >7d → day).
• list_stats returns the account's stats; pass type to filter to counter or value.

Build & run locally

pnpm installpnpm build       # tsc → dist/pnpm start       # node dist/index.js  (stdio MCP server)pnpm test        # vitest run  — fully mocked network, no real HTTPpnpm dev         # tsx watch src/index.ts

pnpm start requires EZSTAT_API_KEY to be set; the server exits with a clear error otherwise (and never logs the key).

Security

• The API key is read only from EZSTAT_API_KEY — never hardcoded, never printed.
• All error messages are scrubbed: the API key is never included in tool results.
• HTTP errors carry a stable machine-readable code (unauthorized, rate_limited, timeout, not_found, server_error, ...) and a human-readable message safe to relay.
• Tests mock the network — no real HTTP leaves the test process.

Registry metadata

For Smithery / mcp.so / Glama listings:

• Name: ezstat-mcp-server
• Display name: EzStat
• Description: Push and read metrics with EzStat from any AI agent. Track counters and values, then ask natural-language questions over your production telemetry.
• Homepage: https://ezstat.dev
• Version: 0.8.0
• Transport: stdio (local) + streamable-http (hosted at https://mcp.ezstat.dev)
• License: MIT
• Repository: this package's source

The mcp block in package.json carries the same metadata so registry crawlers pick it up.

Listing steps (what to submit, where)

1. Smithery: push the repo to GitHub, then run npx -y @smithery/cli publish (or submit via the web UI). The shipped smithery.yaml provides the install schema.
2. mcp.so: submit at https://mcp.so/submit with the metadata above + the GitHub URL.
3. Glama: submit at https://glama.ai/mcp/submit.

The package is not published to npm automatically — publish manually after a release review.

License

MIT.