Use GribStream from ChatGPT, Claude, and Gemini with MCP

Good weather questions rarely start as API requests.

They start as something closer to:

What will the wind and humidity look like in Lisbon tomorrow?

or:

Show me what the last three model runs said for those same hours.

The new GribStream MCP connector is meant for that middle step: turning a weather question into the exact GribStream request you would run. You can point ChatGPT, Claude, Gemini CLI, or another MCP-capable client at:

https://gribstream.com/mcp

The connector is read-only. It does not execute /timeseries or /runs, and it does not need a GribStream API token. Instead, it exposes the catalog and request-building tools an AI needs to build a request you can run through the regular API.

That lets the AI help with questions like:

• What datasets does GribStream support, and how do you know?
• How do I get temperature, wind, and humidity for Lisbon tomorrow?
• Can you turn that point query into a 0.5 degree grid over Portugal?
• What would the forecast look like with a model-run cutoff 18 hours ago?
• How do I query the last three model runs for those same valid hours?
• Can you switch the request from GFS to IFS Operational and fix the selectors?

The result is a regular GribStream API request: a curl, JSON body, headers, and an explanation of what changed.

Why this helps

A forecast request can look reasonable and still be wrong.

An AI can understand "10 meter wind speed", but that is not always a single field in the raw model output. Sometimes it is derived from U and V wind components. It can know "2 meter temperature", but still choose the wrong selector for a specific dataset. It can produce valid-looking JSON, but mix up valid times and run times.

The MCP connector exposes the pieces needed to avoid those mistakes:

• dataset metadata
• exact parameter names and selector variations
• shared parameters such as temperature_2m, wind_speed_10m, and relative_humidity_2m
• expression syntax and supported math helpers
• request builders for /timeseries and /runs
• request validation before the data query is executed

This builds on the public catalog endpoints we released earlier this month.

The workflow

The hosted MCP helps the AI build the request. The regular GribStream API still executes it.

For direct data access, you still use:

• POST /api/v2/{dataset}/timeseries for best eligible values by valid time
• POST /api/v2/{dataset}/runs for run-history and forecast-aging queries

If you are using a local AI coding tool that can run shell commands, set your token in the environment:

export GRIBSTREAM_API_TOKEN='YOUR_TOKEN_HERE'

Then the AI can use the MCP to build the request and use the token to run the normal API call.

For web-based tools such as ChatGPT or Claude, the hosted MCP is mainly a request-building and discovery layer. It can produce the request for you to run; the data query is still a normal authenticated call to GribStream.

Chat your way into an analysis

One useful side effect is that weather analysis can become more conversational.

With an MCP-capable AI tool, you can start broad:

What models does GribStream support for global forecasts?

Then narrow the task:

Build a request for temperature, wind speed, and relative humidity in Lisbon tomorrow.

Then change the geometry:

Make it a grid over Portugal at 0.5 degrees.

Then move into backtesting:

What would this have looked like with a model-run cutoff 18 hours ago?

Then switch endpoints:

Give me the last three runs forecasting those same valid hours.

Then compare models:

Now do the same thing with IFS instead of GFS.

From there, a local AI tool with a configured GribStream token can run the requests and analyze the result: compare models, calculate mean absolute error against an analysis dataset, search for thresholds, or summarize where two models disagree most.

That does not replace careful meteorological analysis. It does remove a lot of the mechanical work between the question and the first useful dataset.

Setup

Setup instructions live on the new GribStream for AI tools page.

At a high level:

• ChatGPT: add a custom MCP connector using Streamable HTTP and https://gribstream.com/mcp.
• Claude: add a custom connector with URL https://gribstream.com/mcp.
• Gemini CLI: add https://gribstream.com/mcp as an MCP server in ~/.gemini/settings.json.

The same page also links to the vendor-neutral GribStream skill file, the OpenAPI spec, the Quickstart, and the Expressions guide.

What comes next

The written examples are useful, but this is the kind of feature that is easier to understand when you watch it happen.

A short video walkthrough would be a good next step: connect the MCP, ask an AI to build a request, run it with a token, then iterate into a small model-comparison or error-analysis workflow.

For now, the connector is available at:

https://gribstream.com/mcp

And the AI setup page is here:

• GribStream for AI tools