Analyga MCP Server

analyga

Agent-first analytics for MCP + web traffic — query, instrument and configure from a terminal or an MCP client. No browser, no dashboard.

A dashboard for humans. tga for your agents.

One analytics surface for both halves of modern traffic: the web visits your site still gets, and the agent/MCP calls that increasingly drive it. Analytics tools assume a human staring at a chart — but increasingly the thing that needs the numbers is an agent: it fired the events, it wants to know whether they landed, and it wants five numbers back, not a dashboard to render. analyga is the surface the agent drives itself: one install, and it queries stats and funnels across web and MCP, sends server-side events, and reads its own schema — all as structured JSON, none of it screen-scraped.

Works with: Claude Code · Cursor · Cline · Windsurf · Aider · Codex · any MCP client

One agent, one terminal — query to funnel in 60 seconds. Watch the full demo.

See it in action

No mockups — every clip is a live tga session against a seeded tenant. Click any GIF for the full-res video.

Install

npm install -g analyga

The npm package is analyga; the command is tga.

Quick Start

# No account yet? Create a tenant and its first key — no card, no admin in the loop
tga signup --slug acme --name "Acme" --username you --password <8+ chars>

# Point at a tenant
tga login --key <your-api-key>     # or set ANALYGA_KEY

# Confirm WHICH tenant you are about to write to (shows tenant + key scope for a tga_ key)
tga whoami

# What can this credential see?
tga websites

# Five numbers, not a dashboard
tga stats <websiteId> --from 7d

# Full reference
tga --help

Every command takes --json — the machine surface is a first-class citizen, not an afterthought:

tga stats <websiteId> --from 7d --json

Read the contract, don't guess flags

tga agent-schema --json

That returns every command, argument, flag, type, whether it is read-only, and the full list of error codes — generated from the same table that drives the parser, the help text and the MCP tools, so it cannot drift from what the CLI actually does. If it is not in the schema, it does not exist.

Query — stats, events, funnels

Ask the questions an agent actually asks after it ships: did the events land, and did anyone make it through the flow?

# Traffic and top-line stats for a window
tga stats <websiteId> --from 7d
tga events <websiteId> --from 24h

# Define a funnel once, then run it whenever
tga funnel define <websiteId> --name signup --steps "/,/pricing,/signup"
tga funnel list <websiteId>
tga funnel run <funnelId> --from 30d

# Cross-site rollup for a team
tga network <teamId> --from 7d

All read-only, all --json-able — an agent branches on the numbers instead of parsing a chart.

Identity — pass --id or your agents collapse into one visitor

Anonymous identity is derived from IP and user-agent. Agent traffic shares both — one CLI version means one user-agent, and agents often share an egress IP — so without a stable id every agent behind that IP collapses into a single visitor. It cannot be backfilled.

tga send <websiteId> --name cli_signup --id <stable-user-id> --data '{"source":"cli"}'

Sending events server-side

# Mint a per-website ingest key first (shown ONCE; regenerating rotates it)
tga ingest-key create <websiteId>

tga send <websiteId> --name cli_signup --id <stable-user-id> \
  --ingest-key <per-site-key> --data '{"plan":"pro"}'

tga ingest health <websiteId> --from 24h    # did anything actually land?

tga ingest-key revoke <websiteId>           # kill server-side ingest for this site

A discarded event still answers HTTP 200. The ingest endpoint returns {"beep":"boop"} when it drops an event as a bot. This CLI treats that as a hard failure with a hint rather than reporting success — so an agent never believes a dropped event succeeded.

Server-side ingest uses a separate per-website key (ANALYGA_INGEST_KEY or --ingest-key), scoped to one website so it can only ever write that site's traffic. Mint one with tga ingest-key create <websiteId> and revoke it with tga ingest-key revoke <websiteId> — the plaintext is shown once and only its hash is stored.

Billing — hand a human a checkout URL

An agent authenticates with its tenant key, names a plan, and gets back a Stripe Checkout URL to pass to a human who completes payment:

tga billing checkout --plan pro

The tenant is pinned server-side from the key. The free plan (and any plan without a configured price) is not purchasable, and internal sister-stack tenants are never billed.

MCP Server

Prefer tools over a CLI? analyga ships an MCP server. Point Claude Code (or any MCP client) at it and your agent gets 13 native tools covering stats, events, funnels, websites, identity, server-side ingest (including minting/revoking ingest keys) and billing checkout.

tga mcp-serve        # stdio, for Claude Code / Cursor / Cline / Windsurf

For clients that use a JSON config, pass your key via ANALYGA_KEY. The MCP server runs outside your project directory, so it will not pick up .analyga/config.json:

{
  "mcpServers": {
    "analyga": {
      "command": "tga",
      "args": ["mcp-serve"],
      "env": { "ANALYGA_KEY": "your-key" }
    }
  }
}

Read-only tools are flagged, so a client will not prompt a human to confirm a read. Every tool declares an outputSchema, so an agent never parses prose or guesses field names — and a tool result and --json output are the same bytes, one client across two surfaces. A session with no key sees only the onboarding surface — how to connect, plus analyga_signup to bootstrap its own tenant — rather than the full toolset.

Error codes

Failures carry a stable code and a distinct exit code, so an agent can branch on the outcome instead of string-matching a message.

Exit 1 is reserved for unhandled crashes, so a classified failure is never ambiguous. Where the server can say what to do next, it comes back as a hint on its own line.

Configuration

Resolved highest-priority first:

1. --key flag
2. ANALYGA_KEY environment variable
3. .analyga/config.json in the current directory

tga whoami reports which of these won. On a multi-tenant platform, "which tenant am I about to write to" is the question that matters most. .analyga/ holds a live credential — keep it out of version control.

Features

• Self-serve signup — tga signup creates a tenant and its first API key with no card and no admin in the loop; keyless, so it works before you have a credential
• Tenant-aware identity — tga whoami shows the pinned tenant and key scope for a tga_ key, or the user and teams for a session login
• Stats & events — top-line traffic and event counts over any window, all read-only, all --json
• Funnels — define once, run anytime; step-by-step conversion for a site
• Network rollup — cross-site stats for a whole team
• Server-side ingest — tga send with a scoped per-site key; mint/rotate/revoke keys with tga ingest-key create|revoke; dropped-as-bot events fail loudly instead of faking a 200
• Billing checkout — tga billing checkout --plan <id> returns a Stripe URL to hand to a human
• Stable identity — --id keeps agents from collapsing into one anonymous visitor
• agent-schema — one command emits the full contract (commands, flags, types, error codes) so agents never guess
• MCP server — 13 tools (tga mcp-serve) for Claude Code, Cursor, Cline, Windsurf, any MCP client
• Classified errors — stable code + distinct exit code per failure class, with actionable hints

Agent Integration

Add to your CLAUDE.md, .cursorrules, .clinerules, .windsurfrules, or AGENTS.md:

## Analyga
This project uses Analyga for analytics. Use the `tga` CLI.
Config is in .analyga/config.json (auto-loaded), or set ANALYGA_KEY.
If not configured: tga login --key YOUR_KEY

Run `tga agent-schema --json` — it returns every command + valid flags + error codes.
This is the single source of truth: if it is not in the schema, do not use it.
Always pass --id on `tga send`, or agent traffic collapses into one visitor.

Why this exists

Every analytics tool I reached for assumed a human looking at a chart — but the thing that needed the numbers was the agent that fired the events. So the CLI and the MCP server are the primary interface here, not a wrapper bolted onto a REST API. It's early and iterating fast: if something's rough or missing, open an issue.

Documentation

For now, tga --help and tga agent-schema --json are the authoritative reference.

Status

Early. Published for dogfooding by the Tyga sister stacks. Interfaces may change; error codes will not be reworded once published, because agents branch on them.

License

Proprietary — Tyga.Cloud Ltd. See LICENSE.