👁 Mata — see your AI footprint

An MCP server that estimates the environmental footprint of your AI use — energy (kWh), miles driven in a gas car, water used for cooling, and CO₂ — plus a prompt-efficiency coach. It works with any AI client (Claude, Codex, others) because it measures token usage, not a specific vendor.

Mata is Tagalog for eye — the tool that lets you see what's normally invisible.

The impact math is a TypeScript port of the EcoLogits life-cycle methodology (CC BY-SA 4.0). See METHODOLOGY.md for sources and attribution.

What we're building (explained simply)

The problem: Every time you talk to an AI like Claude, a powerful computer in a big warehouse (a "data center") does the thinking. That computer burns electricity. Making electricity usually creates pollution (carbon), and the computer gets hot, so the data center uses water to cool it down. Most people have no idea how much — it's invisible.

Mata makes it visible. Like a fitness tracker for your AI use: it counts the AI's words and turns them into:

• ⚡ Energy used (in kWh, like your electric bill)

• 🚗 Miles driven in a gas car that would pollute the same amount

• 💧 Water used to cool the computers

• 🔤 Words in and out (tokens)

Bonus — a prompt coach. If you ask the AI to "make a thing" then correct it five times, that wasted energy. Mata scores how efficiently you set up your work (0–100) and gives tips to get more done in fewer prompts.

Why it's different: it doesn't care which AI you use — it watches the words, so it works across Claude, Codex, and more. Everything stays on your machine, and it stores word-counts, never your chats.

Related MCP server: nikhilnt

The tech stack (explained simply)

A kitchen that turns raw ingredients (AI word-counts) into a finished meal (the impact numbers):

Runs without AI (overhead ≈ 0)

A tool that measures AI's footprint shouldn't burn AI to do it. Mata's entire pipeline — capturing usage, doing the math, scoring efficiency, drawing the dashboard — is deterministic. Nothing in the data path calls an LLM.

Runtime dependencies are just @modelcontextprotocol/sdk, gpt-tokenizer, and zod — none of which make AI or network calls.

Run it with zero AI in the loop. The three standalone binaries need no LLM host:

ai-impact-tail        # watch Claude Code logs and record usage
ai-impact-proxy       # capture exact usage from API clients
ai-impact-dashboard   # render the HTML dashboard

One honest caveat: if you talk to Mata through an AI assistant ("show me my report"), that chat turn is the host's LLM call — Mata's own tools add nothing. Fittingly, Mata would measure that turn too. 🙂

Installation

Requires Node ≥ 22.5 (uses the built-in node:sqlite — no native build step).

git clone https://github.com/anchetadev/mata.git
cd mata
npm install
npm run build
npm test          # 35 tests

Add to an MCP host

Claude Desktop (claude_desktop_config.json) or Claude Code (.mcp.json / claude mcp add):

{
  "mcpServers": {
    "mata": { "command": "node", "args": ["/absolute/path/to/mata/dist/server.js"] }
  }
}

The same dist/server.js works in any MCP-compatible host (Cursor, Cline, etc.).

Optional collectors (run alongside)

node dist/tail-server.js       # watch Claude Code logs, record EXACT usage live
node dist/proxy-server.js      # local proxy: capture EXACT usage from API clients
node dist/dashboard-server.js  # write the HTML dashboard

For the proxy, point a client at it:

ANTHROPIC_BASE_URL=http://localhost:8788   # Claude Code, Anthropic SDK
OPENAI_BASE_URL=http://localhost:8788/v1   # Codex, OpenAI SDK

Usage

Once connected, ask your host things like "estimate the impact of a 2000-token reply from claude-opus-4-5" or "show my AI impact report for this week."

Tools

Resource: impact://methodology — how the numbers are derived.

CLI binaries

How invasive is it?

Short answer: as little as possible, and entirely on your machine. Mata stores token counts and metadata — never the content of your messages — in a local SQLite file at ~/.ai-impact/usage.db. No telemetry, no network calls except the proxy forwarding to the provider you chose.

Each capture method is opt-in and differs in what it can see:

What it cannot see: the Claude desktop/web apps talk to Anthropic directly, so Mata can't passively read them — the web estimator only sees what you choose to hand it. The proxy only sees clients you explicitly repoint.

Every record is tagged exact or estimated, and estimates use a public BPE tokenizer as a stand-in (Claude's isn't public), so they carry ~±10% error.

Accuracy & honesty

These are order-of-magnitude estimates, good for "which of my habits cost the most" — not carbon accounting. Closed-model parameters are inferred; consumer surfaces are re-tokenized. The tool shows its assumptions in every result and ships full sources in METHODOLOGY.md.

License

Dual-licensed (see LICENSE and NOTICE):

• Original code (server, collectors, scorer, dashboard) — MIT.

• Impact engine + vendored data (ported from EcoLogits) — CC BY-SA 4.0 (attribution + share-alike).

Publishing guidance for your own fork: PUBLISHING.md.