Overview

Global weather from Open-Meteo: forecasts, history back to 1940, marine conditions, air quality, probabilistic ensembles, river discharge, and climate projections. Look up any place by name, pull hourly and daily variables, and run SQL over large results. No API key required.

Tools

The server also publishes an agent skill, open-meteo-guide (skill://open-meteo-guide/SKILL.md, source in skills/open-meteo-guide/): which tool answers which question, plus the gates for exceedance odds, frost, comparisons to a normal, climate shifts, surf spots, river cells, and AQI timing.

Capability reference

openmeteo_search_locations _tool

• A bare place name ("Paris", not "Paris, France"), with optional country (ISO 3166-1 alpha-2: two letters, any case, never a country name) and language (default en); count 1–10, default 5
• Each result carries coordinates, an IANA timezone, country / country_code, admin1 / admin2, population, and feature_code; no match fails as no_results
• A notice flags a top match with null or sub-100,000 population, which often means a historic exonym ("Bangalore", "Calcutta")
• A native-script query that misses under en is retried once in the language its script implies; effective_language names it when the retry succeeds (上海 → zh), and results come back localized

openmeteo_get_forecast _tool

• At least one of current_variables, hourly_variables, or daily_variables (up to 50 each), each also accepted as one comma-joined string and under Open-Meteo's own keys current, hourly, daily; forecast_days 1–16 (default 7), past_days 0–92 (default 0)
• cell_selection (land, sea, nearest) picks the grid cell that answers: omitted, Open-Meteo answers from a land cell, so a point offshore reads on-shore conditions. Set sea for a point on the water (wind for sailing, surf, or offshore wind); a point on land keeps the default. The response latitude / longitude name the cell that answered
• Per-timestamp hourly / daily records with hourly_units / daily_units; current_variables adds a current object from 15-minute data (interval 900) plus current_units
• Omitting models reads Open-Meteo best_match; models compares deterministic models (e.g. ncep_gfs_seamless, ecmwf_ifs025, dwd_icon_seamless; older spellings like gfs_seamless still resolve). With 2+ models each hourly and daily column carries a model-name suffix, and models is echoed. current_variables takes at most one model (current_multi_model)
• A regional model with no data at the point is dropped from a multi-model response: the notice names it, or, when one model is left with unsuffixed columns, says the values cannot be attributed to a model; when every requested model misses, the call fails as coverage_gap
• Prefer past_days over openmeteo_get_historical for the last week, where the archive's ERA5 components lag

openmeteo_get_historical _tool

• start_date and end_date required (YYYY-MM-DD, from 1940-01-01), plus at least one of hourly_variables / daily_variables, named as on the forecast tool (a list or one comma-joined string, also under hourly / daily)
• cell_selection (land, sea, nearest) works as on the forecast tool: omitted, a land cell answers, so a point offshore reads on-shore conditions; sea moves it onto the water
• Omitting models reads Best Match (IFS HRES + ERA5 + ERA5-Land, source varies by date); up to 8 models pin a source. The ERA5 family runs about a week behind, ecmwf_ifs has no delay, and cerra covers Europe only
• Output echoes models and the returned date_range; with 2+ models each column carries a model-name suffix. A model with no data at the point (cerra outside Europe) is dropped from a multi-model response; the notice names it, or, when one model is left with unsuffixed columns (era5,cerra), says the values cannot be attributed to a model

openmeteo_get_marine _tool

• One window per call: forecast_days 1–8 (upstream default 7) with past_days 0–92, or a start_date / end_date archive range back to at least 2022. Mixing them fails as forecast_window_conflict, a lone date as date_range_incomplete, and a date outside the range upstream accepts as date_out_of_range
• At least one of current_variables, hourly_variables, or daily_variables (wave height, period, and direction, swell, sea-surface temperature); sheltered water returns near-zero waves, a point well inland returns nulls (named in the notice), a land point near the coast is answered from a nearby sea cell with real values and no notice (check the response's latitude / longitude is the water you meant), and ocean_current_velocity is null off the open ocean
• current_variables takes the hourly names and adds a current object from 15-minute data (interval 900) plus current_units, alongside either window; a daily name there fails as variable_wrong_cadence
• length_unit (metric / imperial) sets every height, sea_level_height_msl included, in m or ft; wind_speed_unit (kmh, mph, ms, kn) sets ocean_current_velocity; temperature_unit (celsius / fahrenheit) sets sea_surface_temperature. Open-Meteo converts the values, and the *_units maps label them (mph reads mp/h)

openmeteo_get_air_quality _tool

• One window per call: forecast_days 1–7 (upstream default 5) with past_days 0–92, or a start_date / end_date archive range from August 2022 (us_aqi starts a day later). Same forecast_window_conflict, date_range_incomplete, and date_out_of_range rejections as the marine tool
• At least one of current_variables (a current object, interval 3600) or hourly_variables: PM2.5, PM10, NO2, SO2, O3, CO, dust, pollen, European and US AQI. Pollen is modeled for Europe only, in pollen season; elsewhere it comes back null and the notice names it
• data_source: "CAMS" marks every response as modeled grid data, not station measurements

openmeteo_get_elevation _tool

• Parallel latitudes[] / longitudes[] arrays of up to 1,000 pairs; unequal lengths fail as coordinate_count_mismatch, compared on the full arrays
• Samples the first 100 pairs (the Elevation API's per-request limit) in one upstream request; unprocessed_count says how many were left and next_offset is where to resume — resubmit latitudes.slice(next_offset) and longitudes.slice(next_offset) until next_offset is null. A call of 100 or fewer returns 0 / null. Each sampled pair counts as one call against Open-Meteo's per-IP limits (600 a minute), so a seventh 100-pair slice inside one minute comes back rate_limited
• elevations[] in input order, each { latitude, longitude, elevation_m } from the Copernicus DEM (~90 m); elevation_m is null where the DEM has no value (latitude 90), without failing the rest of the call

openmeteo_get_ensemble _tool

• At least one of hourly_variables / daily_variables; forecast_days 1–35 (default 7), past_days 0–92. Most models end within 16 days (ncep_gefs05 runs to about day 34); past that, columns are null and the notice says where data ends
• models takes one ensemble name (e.g. ecmwf_ifs025_ensemble, 51 members; gem_global_ensemble, 21); omitted or blank selects ecmwf_ifs025_ensemble, echoed in model
• cell_selection (land, sea, nearest) works as on the forecast tool: omitted, a land cell answers; the response latitude / longitude name the cell that answered
• Each variable returns per-member columns (temperature_2m_member01, …) across up to 64 members; member_count excludes the control run. A regional model queried outside its coverage fails as a non-retryable input error; switch to a global model

openmeteo_get_flood _tool

• daily_variables required (up to 20), all in m³/s: river_discharge plus the statistics river_discharge_mean / _median / _min / _max / _p25 / _p75, which carry ensemble spread on forecast-run days and repeat river_discharge on reanalysis days
• One mode per call: forecast_days 1–210 (the current GloFAS run reaches about six months; later days are null), or a start_date / end_date range accepted from 1984-01-01, though discharge begins 1997-01-01 and earlier days are null. Mixing them fails as forecast_days_conflict, a lone date as date_range_incomplete, a date outside the range upstream accepts as date_out_of_range
• Omitting models reads seamless_v4: consolidated_v4 reanalysis before 2023-08-05, forecast_v4 runs from then on. models: ["consolidated_v4"] reads the reanalysis from 1997-01-01 through 2025-05-31 (null outside that span, and on forecast_days); forecast_v4 starts 2023-06-01. Output echoes models; with 2+ models each column carries a model-name suffix, and the notice names null spans
• Discharge comes from the GloFAS 0.05° grid cell nearest the coordinate, not reliably the cell on the river you mean; when the value is far below what the river carries, step the coordinate 0.05° across the channel and compare. Null outside GloFAS coverage

openmeteo_get_climate _tool

• start_date / end_date within 1950-01-01 to 2050-12-31 and daily_variables required (the API is daily-only); up to 7 CMIP6 models, e.g. CMCC_CM2_VHR4, MRI_AGCM3_2_S
• With 2+ models each variable is suffixed with the model name; a single or omitted model returns plain names. A variable a model doesn't carry comes back null
• cell_selection (land, sea, nearest) works as on the forecast tool, except that the response latitude / longitude can stay the same while the values change with it, so they do not show which cell answered

openmeteo_dataframe_describe _tool

• Takes the canvas_id a spilled weather tool returned; lists each table's name, kind, row_count, and columns (name, type, nullability), plus the canvas expires_at
• Fails as canvas_not_enabled when canvas is off, or canvas_not_found when the ID is unknown or past its 24-hour sliding TTL

openmeteo_dataframe_query _tool

• canvas_id plus a read-only SELECT against the staged table_name; rows holds the first limit rows (default 25, each whole, values never cut). A larger limit returns at most 10,000 rows (or CANVAS_DEFAULT_ROW_LIMIT when lower) and a notice says so. row_count is the exact total and truncated is true when it exceeds rows; page with OFFSET <rows.length> under an ORDER BY
• A page's size rests on limit and the columns selected, so from a member-suffixed ensemble table select the columns the question needs rather than SELECT *. One response carries at most 8,000,000 characters of row data (the rows written as JSON, a character that needs escaping counted at its escaped length); a larger page is refused as result_too_large, never cut, and the message says how many leading rows fit
• If the count fails or runs past 5 seconds, rows still return, row_count is a lower bound, and count_error names the reason (e.g. count_timeout)
• Fails as canvas_not_enabled, canvas_not_found, missing_table, system_catalog_access (system catalogs are blocked, and so is a DuckDB metadata view's name such as pg_class used as a column alias, CTE name, or table alias), or result_too_large. An error message relayed from the query engine is cut to 2,000 characters, keeping its code and reason

openmeteo_dataframe_drop _tool

• Off by default: listed as disabled until OPENMETEO_DATAFRAME_DROP_ENABLED=true
• canvas_id plus the exact table_name from openmeteo_dataframe_describe; removes that one table or view and returns dropped and the remaining_tables. A name that is not staged returns dropped: false, so a repeated drop is safe
• Fails as canvas_not_enabled, canvas_not_found, or invalid_table_name (not a plain SQL identifier, or a reserved keyword)

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

Open-Meteo-specific:

• No API key for non-commercial use; commercial use requires Open-Meteo's paid API tier
• Columnar API responses reshaped into per-timestamp records with *_units maps, on the same variable names across forecast and archive
• DataCanvas spillover on the seven weather tools: with CANVAS_PROVIDER_TYPE=duckdb (the default) an over-budget result is staged as a DuckDB table (response carries canvas_id, table_name, truncated: true); with none it returns a bounded preview and truncated: true
• models on the forecast, historical, ensemble, climate, and flood tools is not an allowlist, so a model Open-Meteo adds later works without a server update

Agent-friendly output:

• Location-first workflow: openmeteo_search_locations returns the IANA timezone alongside coordinates, ready for any weather tool's timezone input (default auto)
• Variable and model lists take an array or one comma-joined string ("temperature_2m,precipitation"); the *_variables fields also answer to Open-Meteo's own current, hourly, and daily keys where the tool has the field
• Typed rejections: variable_wrong_cadence (names the correct field), request_too_large (names the inputs to narrow), coverage_gap (a coordinate outside the dataset's grid or the model's domain), date_out_of_range (a date outside the range the archive, marine, air-quality, climate, or flood endpoint accepts), and rate_limited; only Too many concurrent requests is retried first. An outage that outlasts three attempts fails as upstream_unavailable, or upstream_timeout when the last attempt timed out
• Request URLs never appear in error data or client-visible logs
• One notice per response for what the data alone won't show: unserved variable names, nulls outside a dataset's coverage, and where spilled rows went

Getting started

Connect to the hosted instance, no install:

claude mcp add --transport http open-meteo-mcp-server https://open-meteo.caseyjhand.com/mcp

Or run it locally:

claude mcp add open-meteo-mcp-server -- npx -y @cyanheads/open-meteo-mcp-server

For other clients, add this to the MCP config file:

{
  "mcpServers": {
    "open-meteo-mcp-server": {
      "command": "npx",
      "args": ["-y", "@cyanheads/open-meteo-mcp-server"]
    }
  }
}

No API key is required for non-commercial use; commercial use requires Open-Meteo's paid API tier. For Docker or a source build, see Running the server.

Configuration

Every variable is optional.

See .env.example for the full list of optional overrides.

Running the server

Local development

Requires Bun v1.4.4 or higher (or Node.js v24+).

• Clone and install:

  git clone https://github.com/cyanheads/open-meteo-mcp-server.git
  cd open-meteo-mcp-server && bun install
  

• Build and run:

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:stdio
  # or
  bun run start:http
  

• Run checks and tests:

  bun run devcheck   # Lint, format, typecheck, security
  bun run test       # Vitest test suite
  bun run lint:mcp   # Validate MCP definitions against spec
  

Docker

docker run --rm -p 3010:3010 ghcr.io/cyanheads/open-meteo-mcp-server:latest

The image defaults to HTTP transport, stateless session mode, and logs to /var/log/open-meteo-mcp-server. To run it over stdio from an MCP client, use "command": "docker" with "args": ["run", "-i", "--rm", "-e", "MCP_TRANSPORT_TYPE=stdio", "ghcr.io/cyanheads/open-meteo-mcp-server:latest"]. To build locally, docker build -t open-meteo-mcp-server .; OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage
• Register new tools in the tools[] array in src/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Attribution

Weather data by Open-Meteo.com, licensed CC BY 4.0.

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

Apache-2.0 — see LICENSE for details.