Skip to content

AI Agents (MCP)

Hydrata publishes a Model Context Protocol (MCP) server (MCP is an open standard that lets AI assistants use other services' tools), so an AI assistant such as Claude or ChatGPT can work in your Hydrata account: set up a project, draft its inputs, build a scenario and follow a run. You stay in charge of the steps that matter, which you do in the browser.

In these pages, the assistant is the app you chat with, such as claude.ai, ChatGPT or Claude Code. When it acts in your Hydrata account, we call it an agent.

hydrata.com only

Agent keys and one-click connect work on hydrata.com. They are not available on Hydrata's other sites.

What an agent can do

You give each agent one of three access levels, each one including the one before:

Level What the agent may do
Read Look at your projects, layers, time series and runs
Draft Also create and change projects, layers, time series and scenarios, and add or delete flood observations. Spends nothing
Launch Also build scenarios, prepare terrain (finish an upload, fetch or convert a terrain), derive merged terrains and IDF tables, and start, retry and cancel runs. These use Hydrata's compute, but only starting a run is charged: building and preparing terrain are free (see Compute balance)

A run an agent starts is charged exactly like a run you start yourself, at the same quote. Reading and drafting are free. A run that costs more than the agent may spend without asking waits for your consent.

Whatever an agent reads from Hydrata passes through the assistant you connected, under that provider's terms. Hydrata records each tool call, reads included, and keeps samples of what the agent sent (see What Hydrata records about agent tool calls). If a project is confidential, do not connect an assistant you would not show it to.

What stays with you

Hydrata enforces these limits, whatever the agent is asked to do. No agent, at any level, can do these things. You do them yourself in the browser:

  • review the inputs it wrote;
  • consent to the cost of a run that is over its spend limit;
  • publish a project;
  • change a project's visibility or projection, or delete it;
  • add, remove or change members and invitations;
  • create or change a publication;
  • upload, replace or delete a friction raster, or write a hydrology network;
  • overwrite an existing layer through the importer;
  • create, change or revoke agent keys.

The loop at a glance

  1. Connect. You connect an AI assistant and choose its access level.
  2. The agent drafts. It creates the project, uploads terrain, writes time series and attaches layers. After each step it gives you a link that opens the map on what it changed.
  3. You review. Inputs the agent wrote are marked as not yet reviewed. Build, Run and Build and Run stay off until an Editor reviews them.
  4. You consent to spend. A run over the key's monthly spend limit waits for your consent. Nothing starts until you press the button.
  5. You publish. A project an agent creates starts unpublished: only you and its members can see it, until an Owner or Manager presses Publish.

Every agent write is recorded, and you can download the record. See Agent activity and controls.

Stop an agent

Open the Agent keys page while signed in. Your keys and your one-click connections are both listed there. Press Revoke beside the one you want to stop, or Revoke all to stop every one. See Revoke a key or a connection.

In this section

  • Connect an AI assistant: one-click connect for claude.ai and ChatGPT, agent keys for Claude Code and scripts, access levels, spend limits and revoking.
  • Working with an agent: the links the agent hands back, input review, run consent, publishing, and the rules the agent follows for units and time series.
  • Agent activity and controls: the activity log, downloads, and what happens when Hydrata pauses agents.
  • Tool reference: all 28 tools, the access level each needs, the refusal codes, and the server details for developers.

For the same workflow in the browser, start with the Getting Started tutorial. For the HTTP interface underneath both, see the REST API.