GreenPick MCP server

A design-time sustainability co-pilot that runs as an MCP server inside your vibe-coding environment (Cursor / Claude Code / Kiro). Ask it where to deploy, when to run a job, or how to right-size — and it answers with an honest carbon + cost + latency tradeoff, in-editor, in seconds. (It responds when you ask; auto-nudging on infra-code changes is on the roadmap.)

Status: working — 20 tools. Live carbon for US (EIA-930), EU (ENTSO-E) and Canada (Hydro-Québec, IESO — no key needed), cross-cloud cost parity (Vantage-refreshed), per-axis R/Y/G ratings, carbon-aware scheduling + day-ahead forecast, GPU-training and LLM-inference carbon, right-sizing, serving-pattern and data-locality/residency checks — all shipped. Offline or without API keys, falls back to clearly-labelled typical values: this month's average for each grid, from a year of real data.

The same 40-GPU-hour training run emits ~0.6 kg CO₂ in Stockholm vs ~17.5 kg in Sydney — a ~28× difference from region choice alone (typical October grids; Sydney on Ember's 2025 national figure, a lower bound for its coal-heavier state grid). Ask GreenPick and it surfaces that (with the cost and latency tradeoff) while you're still in the editor, before you commit.

▶ See the 5-minute demo — three scenarios with real tool output (training placement, the egress backfire, the forecast).

Features

• 🧮 Right-sizing that cuts cost and carbon together — the FinOps win with a carbon twin. The money saved is the reason to run it; the carbon comes free.
• ⏱ Carbon-aware scheduling — run deferrable jobs (ETL, batch, training) when the grid is cleanest, now or day-ahead. A decision you make for every recurring job.
• ⚡ GPU training & LLM inference carbon — estimate a run's footprint and the greener where + when; the size/placement choices that actually move the number.
• 🌍 Greener region picks — ranked by carbon, with the cost + latency tradeoff shown, not hidden.
• 🛡 Backfire guards — flags when a greener move loses to egress, or crosses a data-residency line.
• 🔌 Live US + EU grid data, with a realistic fallback built from a year of real data — no API keys required to start.

Contents

• Why this exists
• What it is / is NOT · How it differs
• Scope
• MCP tools
• Data sources
• Honest limitations
• Quick start
• FAQ
• Contributing
• License

Why this exists

The same web service emits very different CO₂ depending on which region it runs in, because electricity grids differ in carbon intensity. Existing tools either rank regions in the abstract, or measure carbon after you deploy. GreenPick does it at the moment you choose, and tells you what the greener choice costs you in latency and dollars — so you can accept or reject it with eyes open.

What it is / is NOT

• IS: directional "greener choice" guidance for a developer picking a cloud region while building. Guidance, with tradeoffs stated.
• IS NOT: audited carbon accounting, ESG/Scope-2 reporting, or a real-time scheduler. GreenPick does not claim GHG-Protocol or ISO compliance — that's what the paid data vendors (Electricity Maps, WattTime) provide. GreenPick is for decisions, not disclosures.

How it differs from what's out there

• Measurement tools (Cloud Carbon Footprint, CodeCarbon, Kepler) report what you already emitted, after you deploy. GreenPick advises before — at the choice.
• Cloud providers' own tools (AWS CCFT, Azure, Google's "Low CO₂" region labels) cover one cloud and show carbon alone. GreenPick is cross-cloud and weighs carbon against cost and latency so you see the tradeoff, not just a green badge.
• CI/CD carbon tools act at merge time on a pipeline job. GreenPick works earlier, in the editor, across more decisions (scheduling, right-sizing, training/inference, egress & residency).
• Grid-data SDKs (Electricity Maps, WattTime, GSF Carbon Aware SDK) are building blocks; GreenPick is the decision layer on top of that kind of data.

The combination — cross-cloud, carbon + cost + latency, at design time, in the editor, with backfire guards — is the part nothing else puts together.

Scope

• Providers: AWS, GCP, Azure — 25 regions across the US, Europe, Canada, Asia and Australia.
• Workloads: long-running web/API services, GPU training runs, and deferrable jobs (ETL / batch / backups) — placement and timing for each.
• Carbon = grid carbon intensity by region (directional, point-in-time; live US + EU), plus a day-ahead forecast (EU) for forward-looking scheduling.
• Cost = compute + egress ballpark (deterministic static table, apples-to-apples across clouds; never LLM-invented).
• Latency = geographic estimate to a stated user location.
• Efficiency levers: greener instance family (perf-per-watt), right-sizing (cost + carbon together), and a carbon-metric scaffold to track your app over time.
• Guardrails: data-locality / egress backfire checks and cross-jurisdiction residency flags.
• Free data only. No flow-traced consumption accounting; carbon forecast is relative, not absolute.

Roadmap: data-format/engine efficiency advice, sub-national grid zones, CI/CD integration. See CHANGELOG.md for what has shipped in each version.

MCP tools

20 tools, grouped by what you're trying to do.

Pick a region

Tool	What it does
compare_regions_tri_axis	Rank candidate regions by carbon with cost + latency tradeoffs; recommend a pick, each axis rated R/Y/G. Pass regions to rank only the ones you'd actually consider.
greener_alternative_to	"You picked X — the greenest region within +N ms latency is Y (−Z% carbon, Δcost)." The core nudge.
region_carbon_now	Current gCO₂/kWh for a region, with source + timestamp + freshness — and, when live, whether that's cleaner, dirtier or about usual for that hour this month.

Time a job (scheduling)

Tool	What it does
cleanest_hours_to_run	Cleanest recent hours to time a deferrable job (ETL, backups, training), + % reduction. Pass window_hours for a job of known duration → cleanest contiguous window to start at.
cleanest_time_of_day	For a recurring job (nightly batch): the region's cleanest time of day, averaged across recent days. A daily pattern, not a 7-day weekly one.
greenest_regions_on_average	Which regions are cleanest on a typical day in a given month (or the whole year), ranked from last year's real data. Offline, no keys; a planning view, not today's grid.
carbon_forecast	Forward-looking: which upcoming hours will be cleanest, from the EU day-ahead wind+solar forecast (ENTSO-E). Relative ranking; US/uncovered zones degrade to the recent historical pattern.

Estimate a workload's carbon

Tool	What it does
estimate_workload_carbon	Rough kgCO₂/month for a workload in a region (an estimate, not a measurement).
estimate_training_carbon	Carbon of a GPU training run (TDP-based energy × live grid), with greener where (region ranking) + when (cleanest hours) — the two levers that compound for portable, deferrable training.
estimate_inference_carbon	Energy + CO₂ of an LLM inference workload, by model size class × output tokens × requests (× live grid). A low–high band; smaller size class is the big lever. Self-hosted-friendly; not a named-model leaderboard.

Efficiency & guardrails

Tool	What it does
greener_instance_family	A greener same-size family in the same region via perf-per-watt (newer generation / ARM Graviton-class does the same work on less energy).
rightsizing_check	Right-sizing (the #1 FinOps win) with a carbon twin: if you're over-provisioned, a smaller size cuts both cost and carbon → the right-sized target with monthly $ and kg CO₂ saving.
serverless_check	For a mostly idle or spiky service: an always-on server vs. paying per request (AWS Lambda, Cloud Run). Monthly cost of each, the break-even request count, and estimated carbon (serverless as a lower bound).
offhours_check	For dev/test/staging that doesn't need to run 24/7: the monthly $ and kg CO2 saved by running it only N hours a week. Compute only, directional.
data_locality_egress_check	Warn when running compute away from where your data lives backfires: directional egress $ + coarse transfer carbon, and whether a greener-region move is still a net win.
migration_payoff	Does moving the data to the compute's region pay off? Payback in months, one-off vs monthly egress and transfer carbon, net over 12 months. Egress only.
region_move_justification	A short paste-ready paragraph for a pull request or decision record, with the carbon, cost, latency and egress numbers and their quality labels. Says so if the move is not an improvement.
serving_pattern_advice	For a user-facing platform: serve from one region or several? Traffic-weighted latency + carbon, with the multi-region data-sync egress and cross-jurisdiction residency catches.

Instrument & understand

Tool	What it does
carbon_metric_scaffold	Emits paste-in instrumentation (Python) so your app reports its own estimated carbon over time — observable proxies × GreenPick's power model × live grid. Self-tracking, directional; not a runtime meter.
explain_methodology_and_limits	The assumptions and what GreenPick explicitly does not claim.

Data sources (free; swappable behind one interface)

• EIA-930 (US balancing authorities) and ENTSO-E (EU zones) — live grid carbon, auto-routed by region. Both also drive cleanest_hours_to_run from recent hourly history.
• Hydro-Québec (Montreal / Québec City regions) and IESO (Toronto regions) — live hourly generation by source, public, no key needed. Hydro-Québec figures arrive 1.5–6 h late, so the latest complete hour is used and its timestamp shown.
• Typical — the offline default and fallback: for each region, this calendar month's average intensity and average hour-of-day profile (UTC), computed from 12 months of real hourly EIA-930 / ENTSO-E data (data/typical.json, refreshed by scripts/refresh_typical.py). Labelled quality: "typical" — a realistic norm, never presented as a current reading.
• Annual — for Singapore, Tokyo, Mumbai/Pune and Sydney, where no free hourly data exists: the country's yearly lifecycle grid intensity from Ember (via Our World in Data, CC BY 4.0), with source and year (data/annual.json). Labelled quality: "annual_average"; fine for comparing regions, but carries no timing signal.
• Mock — a hand-set illustrative value, labelled quality: "mock", kept only as a last resort if every other source is missing. No catalog region relies on it.
• Cost: a dated static pricing table (data/pricing.json), refreshed from live Vantage pricing via scripts/refresh_pricing.py using one matched method across all three clouds.
• Instance-family efficiency: directional perf-per-watt multipliers (data/instance_families.json).
• Latency: great-circle distance from region coordinates to the user's location.

Set EIA_API_KEY and ENTSOE_TOKEN (both free) for live carbon — see Turn on live data; GREENPICK_CARBON_SOURCE selects the source (auto routes US→EIA, EU→ENTSO-E, falling back to typical).

The data layer is a swappable repository (typical values by default, works offline). Nothing here invents numbers: the LLM writes the rationale, the code supplies the figures.

Honest limitations (shipped in the tool itself)

Directional, not audited. Region→grid-zone mapping is approximate. Carbon is point-in-time, not flow-traced consumption. Cost is a ballpark, not a quote. Latency is a geographic estimate. "Greener" assumes grid intensity dominates and ignores embodied/hardware carbon and provider RECs.

Quick start

# Python 3.11+
pip install greenpick          # from PyPI
python -m greenpick            # run the MCP server (stdio)

Or from source, for development (editable install + tests):

git clone https://github.com/bithug01/greenpick.git
cd greenpick
pip install -e ".[dev]"
cp .env.example .env            # defaults to typical values — works offline
python -m greenpick             # run the MCP server (stdio)
pytest                          # run tests

Connect it to an MCP client (e.g. Kiro / Claude Code / Cursor) by pointing the client at python -m greenpick in this directory. See .env.example for switching the data source.

Turn on live data

GreenPick runs on labelled typical values until you give it two free keys: EIA (US grid) and ENTSO-E (EU grid). Pass them as environment variables. Canadian regions are live with just GREENPICK_CARBON_SOURCE=auto — their sources need no key. (A .env file works for a source checkout; if you installed with pip install greenpick, use one of the options below.)

In an MCP client (Kiro, Claude Desktop, Cursor…), add an env block to the server entry:

{
  "mcpServers": {
    "greenpick": {
      "command": "python",
      "args": ["-m", "greenpick"],
      "env": {
        "GREENPICK_CARBON_SOURCE": "auto",
        "EIA_API_KEY": "your-eia-key",
        "ENTSOE_TOKEN": "your-entsoe-token"
      }
    }
  }
}

In a Jupyter notebook or Python script, connect with the MCP client library and pass the keys in env. The library starts GreenPick as a separate process and doesn't forward the notebook's own environment variables to it:

import sys
# "mcp" is the MCP Python SDK (installed alongside greenpick), not a greenpick module.
from mcp import ClientSession, StdioServerParameters
from mcp.client.stdio import stdio_client

params = StdioServerParameters(
    command=sys.executable,
    args=["-m", "greenpick"],
    env={
        "GREENPICK_CARBON_SOURCE": "auto",
        "EIA_API_KEY": "your-eia-key",
        "ENTSOE_TOKEN": "your-entsoe-token",
    },
)

# Keep errlog=: in Jupyter it must be a real file, because the notebook's stderr has no fileno()
# (without it you get "UnsupportedOperation: fileno").
with open("greenpick-server.log", "w") as errlog:
    async with stdio_client(params, errlog=errlog) as (read, write):
        async with ClientSession(read, write) as session:
            await session.initialize()
            result = await session.call_tool(
                "region_carbon_now", {"provider": "aws", "region": "us-west-2"}
            )
            print(result.content[0].text)

Top-level await works in Jupyter. In a plain script, put the with block inside an async def and run it with asyncio.run(...).

Using an AI assistant inside a notebook? Paste the starter prompt as your first message: it sets up the connection and points the assistant at the right tool for each question.

Every carbon figure includes a quality field: live means a current reading, typical means this month's average from last year's real data, annual_average means a sourced yearly national figure (regions with no free hourly data), and mock means a hand-set value. Check it to confirm live data is on. A result that isn't live also includes a live_data_hint saying why (no key set, the source didn't respond, or no free source covers that region) and what to do about it.

FAQ

Do I need API keys? No — out of the box GreenPick uses typical values: each grid's average for the current month, from a year of real hourly data, labelled quality: "typical". For live carbon, set two free keys (EIA_API_KEY for US, ENTSOE_TOKEN for EU) and GREENPICK_CARBON_SOURCE=auto — see Turn on live data.

Why "directional" and not audited numbers? Because it's for a design-time decision, not a compliance report. "Stockholm is ~28× cleaner than Sydney" is what you need to pick a region in seconds — a certified figure is a different job for a different tool (Electricity Maps, WattTime). GreenPick is for decisions, not disclosures, and says so in every tool's limitations.

My cloud says this region is already 100% renewable — so does region even matter? That's market-based accounting (the provider matches usage with renewable certificates/PPAs over a year). GreenPick uses location-based accounting — the actual carbon intensity of the grid you physically draw from, hour to hour. Both are legitimate (the GHG Protocol asks for both), and they answer different questions: market-based is about procurement claims; location-based is about the real grid your workload leans on right now. GreenPick is deliberately location-based because that's what a placement/timing decision actually affects — you can't change the provider's annual REC portfolio, but you can choose a cleaner grid or a cleaner hour. We ignore RECs by design, not oversight.

What happens for a region with no live data? It falls back to that region's typical value for the month (or, for a region no free source covers, a hand-set mock value), and labels which — it never presents a fallback as a live reading. Forecast likewise degrades to the recent history, or the typical day, where no day-ahead feed exists (e.g. US), labelled as such.

Can it rank Claude vs GPT vs Llama for carbon? No, by design. estimate_inference_carbon works by model size class (self-hosted-friendly), not a named-model leaderboard — vendors don't disclose the serving details that would make such a ranking honest.

How do I ask it something? You don't call tools directly — you talk to your MCP client in plain language ("where's the greenest region to run this?") and it invokes the right tool. See the demo for the kinds of questions that trigger each one.

Contributing

Contributions of any kind are welcome — a bug report, a data-source correction, a sharper caveat, or code. See CONTRIBUTING.md for the short version. The one thing GreenPick asks: keep it honest and directional — no invented numbers, no precision the data can't back (that discipline is the whole point).

License

MIT — see LICENSE. © 2026 Data Science Advisers, LLC.