GageLink

Hydrology data for AI agents, with the reference frames kept attached.

River levels, streamflow, flood forecasts, drainage basins and satellite water levels from USGS, NOAA, Hub'Eau, the UK Environment Agency and SWOT. Every value arrives carrying its unit, the datum it was measured from, its timezone, and whether the record is provisional or approved.

mcp-name: io.github.Adeniyikayodee/gagelink

Pre-alpha. The API may change.

Install

pip install gagelink for the library. From an MCP client, with nothing installed:

{
  "mcpServers": {
    "gagelink": {
      "command": "uvx",
      "args": ["--from", "gagelink", "gagelink-mcp"]
    }
  }
}

Or open the .mcpb bundle, which carries the server and its dependencies in one file.

No account is needed. A free key raises the allowance from 50 requests an hour to 1,000; set it as GAGELINK_API_KEY.

What can it answer?

• How high is the river, and how does that compare with flood stage?
• How much freeboard is there between the water and a surveyed levee crest?
• What is the flow now, and what fraction of the record peak is that?
• What is forecast over the next few days, and does it cross a flood category?
• What lies upstream or downstream along the river network?
• How large is the basin draining to this point?
• What did a station record over a date range, and has that record been revised?
• What is the water surface elevation of a river with no gage on it?
• Is a reading provisional or approved, and how old is it?

Why the frames matter

At Little Falls on the Potomac, a river stage of 3.02 ft is measured upward from the gage's own zero. A surveyed levee crest of 41 ft is measured upward from a national datum. Both are lengths in feet, so subtracting one from the other produces a number that reads as freeboard, and a units library will pass it.

The gage zero at this station sits 37.04 ft above NAVD88, so the stage is 40.06 ft on that datum and the freeboard is 0.94 ft. Subtracting without the offset gives 37.98 ft, which overstates the margin by a factor of 40 in the direction of calling a levee safe.

GageLink refuses that subtraction and returns the offset that makes it well defined. The same applies to satellite elevations, which sit on a geoid, and to modelled flows, which may have no measurement behind them.

python demo/freeboard.py runs the whole example offline from recorded responses.

Converting a datum

The offset is available for most stations, so the refusal can become an answer. Pass on_datum to describe_location and the station's offset is converted through NOAA's VDatum, with the uncertainty of the conversion returned beside it:

altitude_of_gage_datum        4860 ft (NGVD29)      Boulder Creek at mouth, CO
altitude_accuracy             10 ft, interpolated from a topographic map
altitude_on_requested_datum   4863.061 ft (NAVD88)
conversion_uncertainty        0.17 ft
offset_uncertainty            10 ft

Two things this surfaces are easy to miss.

The offset has an accuracy of its own. Across 7,361 USGS stream stations sampled in four states, 3,397 publish an altitude for their gage datum. Of those, 72% are known no better than a foot. The commonest published accuracy is 15 ft, a third were interpolated from a topographic map, and about one in twenty is levelled to a hundredth. A freeboard is bounded by that figure whatever precision the stage was read to, so describe_location returns it alongside the method used to determine it.

Most stations are on the older datum. 58% of those altitudes are published on NGVD29 while a modern survey or lidar product is on NAVD88. Across the contiguous states the difference runs to feet.

on_datum also takes the tidal datums (MLLW, MLW, LMSL, MTL, DTL, MHW, MHHW) for questions about level relative to the tide, and get_satellite_passes takes it to move SWOT elevations off the EGM2008 geoid they are measured against. Both cover the contiguous United States. Outside that coverage the conversion is refused and the reason is stated.

Tools

All thirteen are read-only and annotated as such, so a client asks for consent once.

Results come back as structured data against each tool's output schema, so a unit, datum or grade is a field the client can read directly.

A series is returned as a handle with a summary. A year of 15-minute record is 35,000 values, and no answer needs them in a context window.

Prompts

A tool list says what can be called. It does not say what to call first, and the datum rule above is an order of operations rather than a call. Four prompts state the ones that go wrong when a model assembles them itself.

Resources

The tables that do not change between calls, readable without spending a request against the hourly allowance.

gagelink://parameter/{code} and gagelink://datum/{name} are templates, and their variables complete: the server answers completion/complete for them and for the prompt arguments with a closed set of values. The specification scopes completion to prompt arguments and resource template variables, which is why those two vocabularies are exposed as templates as well as inside the tool schemas.

Coverage

ERA5, GRACE, CAMELS and HydroBASINS are available to library callers.

Each service publishes a different amount, and the tools say which. Hub'Eau states no unit on any value, so levels in millimetres and flows in litres per second are labelled here from a recorded table. The Environment Agency publishes no record grade on live data, so age is the only staleness signal for a UK reading.

To find a UK station, find_locations takes country=GB. The agency matches river and town in full and in its own spelling, so River Thames returns stations and Thames returns none. Free text matched against the station name is the filter to use when the agency's spelling is unknown.

Protocol support

GageLink serves MCP revision 2026-07-28 and the three handshake revisions before it (2025-06-18, 2025-03-26, 2024-11-05). It declares tools, prompts, resources and completions, and declares listChanged false on all of them: every list is built at import, so a client that subscribed would be waiting on a notification that cannot come.

The 2026 revision removed the initialize handshake. Every request carries its own version and capabilities, so a client calls a tool on its first message and learns what the server is through server/discover. Clients on the earlier revisions continue to open a session and keep it.

Because a connection no longer implies a conversation, a client that wants a ledger of its own names one in _meta:

{"_meta": {"io.github.adeniyikayodee.gagelink/conversation": "whatever-you-call-it"}}

Each name gets its own manifest, quantities and checks. A client that sends no name shares the default.

For clients that cannot start a local process:

gagelink-mcp --http          # http://127.0.0.1:8765/mcp

This binds to loopback and checks the Origin header. It has no authentication, so --host on a reachable interface gives away your hourly allowance.

Reproducible answers

Every retrieval is recorded with its URL, the time it was made, and a hash of the response body. export_manifest returns that record, and a session can be replayed later in three modes:

• offline uses the archived bodies
• strict checks the live service returns identical data
• revision_aware separates a changed answer caused by an official record revision from one caused by changed code

The third mode exists because hydrology data is revised. A provisional measurement is often approved or corrected months later, so an answer can change for reasons that have nothing to do with the code. revision_aware tells the two causes apart.

Values are also checked against the ledger, so an answer can be audited:

[ok]         3.02 ft        from get_latest.00065
[ok]         2960 ft3/s     from get_latest.00060
[UNSOURCED]  116000 ft3/s   no tool output produced this value

Benchmark

waterbench measures whether the interface changes what a model gets right. It runs the same nine tasks under three conditions: raw API responses, structured results with the metadata stripped, and the full toolkit.

First results, gpt-oss-120b, eight replicates, 216 runs:

Six of the nine tasks sit at ceiling, which is a finding about the suite. Where it separates, the causes are legible. Two long-record tasks sent 49,864 and 42,006 prompt tokens through raw JSON against 5,462 and 2,384 through the toolkit. On the opaque-unit task, stripping the reference frames sent seven of eight runs into the recorded trap, answering with the USGS discharge of 3010 ft³/s where the forecast service had published 2.95 kcfs.

One model and a small suite, so these numbers are an early signal about the interface. A general claim would need more models and more tasks.

Development

python3 -m venv .venv
.venv/bin/pip install -e ".[dev]"
.venv/bin/pytest

Requires Python 3.10 or later. The suite answers from recorded fixtures and needs no network access. mypy src/gagelink is expected to be clean.

License

MIT