Baseline Mcp

Baseline Climate

Stop your agent inventing climate context. Ranked 35-year records for any point on Earth, cross-checked against real stations, with an explicit reliability note to relay instead of guess.

Ask it something, and this is what comes back.

A ranking — with the near-tie disclosed, and a station agreeing:

August 2026: Mean temperature was 75°F, +2.8°F from normal, ranking as the 8th warmest since 1991 — though the 7th warmest through 9th warmest years are within 0.7°F of each other. A nearby station (CASPER-NATRONA COUNTY AP, 8.4 mi) agrees: +2.1°F from its own normal, against the reanalysis's +2.8°F.

A station that does not agree — the gauge a mile away read half the grid's total:

March 2026: Precipitation totaled 2.30 in — near average for this period historically. A nearby station (BOZEMAN MONTANA STATE UNIVERSITY, 1.3 mi) measured 1.12 in, against the reanalysis's 2.30 in.

A wind ranking that says plainly it is modeled, not measured:

March 2026: Modeled peak gust: 60 mph, ranking as the 3rd gustiest since 1991 — though it's within 1 mph of the 6th gustiest year. A station (Saratoga, 32.1 mi) measured a peak gust of 59 mph. The reanalysis figure is a modeled estimate, not a direct reading.

A seasonal outlook carrying its own skill label:

Leaning wetter than normal for Sep–Nov in Nairobi, on odds we haven't been able to verify here — below normal 5%, near normal 19%, above normal 76%. Forecasts here do track wetter and drier seasons, so the direction is the part worth using; whether the percentages themselves beat a climatological guess has not been shown.

Four things worth noticing. It gives you the rank, not just the number. It tells you when the rank is a near-tie, so you don't over-read it. It goes and asks a real weather station whether the gridded record is right — and when the station disagrees, it says so rather than papering over it. And when it can't vouch for its own forecast, it tells you that too.

The data isn't new. Asking it a question is.

ERA5 has been sitting in public archives for years. Interrogating it meant knowing which dataset, writing the code, and having an afternoon. That cost is what collapsed — and the interesting consequence is that people now ask things they'd never have thought worth the trouble.

The Bozeman example above isn't a weather lookup. It's a construction schedule dispute, where the question was whether rain actually stopped work — and the gridded record said one thing while the gauge a mile away said another. The wind example is a trucking corridor. Neither would have been worth opening a climate dataset for. Both take a sentence now.

That's the claim: not that historical weather data exists, but that the cost of asking it a real question has fallen far enough that weather becomes one ordinary input into questions that aren't about weather.

What it won't claim

Most tools hand you a number. This one hands you the number and its standing.

• Near-ties are disclosed. "8th warmest — though the 7th through 9th are within 0.7°F" stops a model reporting a rank as though it were a meaningful gap.
• Station disagreement moves confidence, and says which way. A gauge reading 1.12 in against a grid's 2.30 in is reported, not averaged away.
• When no cross-check ran, it says why — no nearby station, outside the covered region, source didn't respond. The gap is named rather than left for the model to fill.
• Seasonal outlooks carry an honest skill label. Calibrated from ECMWF SEAS5 against 35 years of observations, then labelled by what testing actually showed: that the odds beat a climatological guess here, or that only the direction is worth using, or — where a simple warming trend predicted the season better than the model did — that the model isn't the thing to go on. We publish the cases where our own forecast isn't the best available answer.

Every response carries an assessment field with Baseline Climate's own confidence statement, and the tool descriptions instruct the model to relay it rather than compose its own. That exists because the failure mode is real: given a bare number, models invent records that weren't broken, assert a dataset is unreliable somewhere it isn't, and recycle one station disagreement across unrelated answers.

What it's made of

Rankings use the full 1991–present record; "normal" means the WMO 1991–2020 reference period. See METHODOLOGY.md for how rankings, percentiles and water years are computed, and where the limits are.

Tools

• get_climate_context — a natural-language question; forecast plus 35-year ranked context.
• get_context_for_coordinates — the same for an exact lat/lon.
• get_water_year_status — precipitation and temperature since the start of the water year, ranked against 35 years.
• compare_to_normal — how unusual near-term conditions are at one location.
• compare_locations — rank precipitation, temperature or snowfall across 2–10 locations, or a curated category, over a water year, season, month or ski season.

Install

Requires an API key. Self-serve signup isn't available yet — email Chad McNutt ([email protected]) during this early period.

pip install baseline-mcp     # or: uvx baseline-mcp

Claude Code

claude mcp add baseline \  --env BASELINE_API_URL=https://api.baselinecontext.com \  --env BASELINE_API_KEY=your-key-here \  -- uvx baseline-mcp

Claude Desktop — ~/Library/Application Support/Claude/claude_desktop_config.json:

{  "mcpServers": {    "baseline": {      "command": "uvx",      "args": ["baseline-mcp"],      "env": {        "BASELINE_API_URL": "https://api.baselinecontext.com",        "BASELINE_API_KEY": "your-key-here"      }    }  }}

Cursor — .cursor/mcp.json, same shape as above.

As an HTTP server (streamable HTTP, for hosting behind a reverse proxy):

BASELINE_API_URL=https://api.baselinecontext.com baseline-mcp --transport http --port 8765

It serves /mcp on 127.0.0.1 and holds no key of its own: every request must carry the caller's key in an X-Api-Key (or Authorization: Bearer) header, and each tool call is made with that key, so it counts against that key's own quota. A request without a valid key gets a 401. The default, --transport stdio, is the local server shown above.

Contributing

This package is a thin translation layer: no climate logic lives here, and every tool call is an HTTP request to the Baseline Climate API. See CONTRIBUTING.md in the repository for local development setup.

MIT licensed.