Open-source · Apache-2.0 · Agent-native

Utility bills,
built for agents.

Retrieve, normalize, review, and export utility bills from fragmented provider portals — one stable core, installable provider plugins, and an MCP server your AI agents connect to directly. Governed, audited, evidence-first.

MCP · https://utilitywatch.kapitec.pro/mcp
Utility Watch logo
MCP server · live & agent-ready

The problem

Utility cost is real financial data — locked behind dozens of provider portals, each with its own login, layout, anti-bot defenses, and PDF format. One-off scrapers rot the moment a portal changes.

1 · Install

Add a utility company as an installable plugin. Each plugin knows one portal; the core stays stable.

2 · Retrieve

An agent or operator runs a retrieval. Hard portals escalate to the Bright Data Scraping Browser, opt-in and budgeted.

3 · Review & export

Every bill is normalized with a confidence score and evidence, routed to review, then exported as clean JSON.

One platform, three faces

For agents — MCP

Connect Claude Code, ChatGPT, or any MCP client. Tools to list providers, run retrievals, query bills, and diagnose runs.

/mcp

For humans — Console

A management console: providers, accounts, bills, run history, users, and reports. Sign in to act.

Open console →

For systems — API + CLI

A service layer behind every surface, plus a CLI for setup, doctor, and operations. No logic duplicated.

How an agent uses it

Register the MCP endpoint once, then your agent discovers and calls tools directly — list providers, run retrievals, and query normalized bills with per-field confidence. No glue code.

Read the tool reference →
# 1 · register the MCP server with your agent $ claude mcp add --transport http utility-watch https://utilitywatch.kapitec.pro/mcp # 2 · the agent calls a tool > list_bills { "provider": "mock-power", "limit": 2 } [ { "id": "bill_8f2", "amount": 142.50, "currency": "USD", "confidence": 0.98 }, { "id": "bill_8f3", "amount": 137.10, "currency": "USD", "confidence": 0.95 } ]

Agent-native, not agent-omnipotent

Agents and operators carry capability-scoped tokens. Reading and running retrievals is open; finalizing exports, approving bills, or spending Bright Data budget is fail-closed — it needs a human or an explicit grant. Every action is audited.

Per-field confidence + evidence on every bill
Human review gate before export
Full audit trail per run
Synthetic fixtures — no private data in the open repo
Bright Data Scraping Browser, opt-in & budgeted

Stop maintaining scrapers. Start installing providers.

Open the console Star on GitHub

Open-source under Apache-2.0 · Built for the Bright Data "Web Data UNLOCKED" hackathon · Finance & Market Intelligence