Hydrology and hydraulics your AI agent can run
Connect Claude, ChatGPT, Cursor or another MCP client to Delta Hydro Studio. The agent runs the calculation engines on your account and saves every result for you to review.
80 tools · streamable HTTP · OAuth or API key
Your agent runs the calculations. You review the engineering.
Delta Hydro Studio has an MCP server: the calculation engines behind the calculators, published as 80 tools an AI agent can call. Ask your agent for a design flood or a culvert and it works down the chain on those engines. Every result lands in your project as a record you can open on the map, check and export.
Size a box culvert for the 1:50 flood at the crossing I have selected in QGIS. Tailwater 0.9 m.
- QGIS Read the selected crossingRoad 103.20 m · inverts 100.00 / 99.76 m · 24 m long
- Studio
delineate_watershedCatchment 4.2 km² - Studio
compute_time_of_concentrationTc 0.62 h - Studio
compute_design_rainfall1:50 · 51.7 mm/h - Studio
run_rationalQ 1:50 = 21.1 m³/s - Studio
size_culvert3.6 × 2.4 m box · HW/D 1.07
Five records are saved in the project. Two things are yours to decide: the runoff coefficient, which I took as 0.35 from the published table, and outlet protection, because the barrel discharges at 6.1 m/s.
-
Nothing to type in
The agent passes the inputs to the engine itself, including what it reads from your other tools: a point from QGIS, levels from a spreadsheet. No forms, and no copying values from one program to the next.
-
Engines do the arithmetic
Every number comes from the engines behind the Studio calculators: FHWA HDS-5 for a culvert, the Rational method or TR-55 for a peak. The language model never computes a result.
-
A record to check and sign
Each run is saved to your project with its inputs, marked as agent work in the audit trail, and opens in Studio with charts, tables and Word and Excel exports.
Bring your own agent. It brings the rest of your tools.
Studio is built for the agent you already use, not for a chatbot of its own. Claude, ChatGPT or Cursor is connected to more of your work than an assistant inside our app could be: your GIS, your files, your email. Add Studio to its connectors and it carries the data across. It reads a crossing in QGIS, runs the hydrology and hydraulics in Studio, and writes the result where you need it.
- QGIS
Layers, selected features and survey points, through a QGIS MCP server.
- Files and spreadsheets
Level schedules, survey exports and earlier reports.
- Documents and email
The brief, the client’s questions, the memo the results go into.
- 80 calculation tools
- Records saved to your project
- Word and Excel reports
From a layer of crossings to a flood study per site
The agent reads the outlets from your QGIS layer and submits them to Studio as one batch, quoted before it runs. Flood studies cover South African sites.
From a level schedule to sized culverts
It takes the road and invert levels from your spreadsheet and sizes each crossing on the HDS-5 engine, with a saved record for every culvert.
From saved records to the design memo
It reads the results back from Studio and drafts the memo in your own document tool, with a link to each record.
Connect, ask, review
Connect your agent
Add Delta Hydro Studio to the agent you already use. In Claude it is a connector you approve once; in Claude Code it is one command; ChatGPT and Cursor take the same address.
Ask for the work
Describe the job the way you would brief a graduate: the site, the return period, the levels you have. The agent chooses the tools, runs them in order and files each result in your project.
Review and decide
Open the records in Studio. Check the catchment, the inputs and the result, change what you disagree with, and export the report when it is right.
Settings → Connectors → Add custom connector, then paste the server address. No API key: you approve access once, on a consent screen in Studio.
https://studio.deltahydro.tech/api/mcp Create an API key in Studio under Settings → API keys, then register the server. Cursor and VS Code take the same address and header.
claude mcp add --transport http delta-hydro \
https://studio.deltahydro.tech/api/mcp \
--header "Authorization: Bearer dh_your_key_here" The set-up guide has Claude, Claude Code, Cursor and VS Code step by step, the check command and troubleshooting. ChatGPT takes the same server address in its connector settings.
80 tools, each one a calculation engine
The tools take the same inputs as the calculator pages and run the same code. An agent can chain them, from a pin on the map to a sized structure, and each saved record comes back with a link into Studio.
Catchment
Snap an outlet to the channel, delineate the catchment, cut the watercourse profile and work out the time of concentration.
- snap_to_channel
- delineate_watershed
- run_watercourse_profile
- compute_time_of_concentration
Rainfall
Design rainfall for every duration and return period from one call, IDF fits, areal reduction, design storms and PMP.
- compute_design_rainfall
- fit_idf_curves
- compute_arf
- run_design_storm
- compute_pmp
Peak flow
Rational and TR-55 peaks and hydrographs, with the published runoff coefficient and curve number tables to cite.
- run_rational
- run_tr55
- generate_hydrograph
- lookup_runoff_coefficient
- lookup_curve_number
Flood studies
A full Design Flood Estimation for one outlet or a batch of up to 200, quoted before it runs, plus the regional methods.
- quote_flood_estimate
- submit_flood_estimate
- submit_flood_batch
- get_flood_request
- run_rmf
Hydraulics
Size or check a culvert, analyse a channel and its lining, run a backwater profile, design a weir, route a detention basin, size riprap.
- size_culvert
- design_culvert
- analyze_channel
- compute_gvf_profile
- design_detention_basin
- design_riprap
Gauges and frequency
Find stream gauges near a site, fetch a gauge’s annual peaks, fit the flood frequency curve and attach it to the flood study it belongs to.
- search_stream_gauges
- run_flood_frequency_analysis
- link_analysis_to_flood_request
Dams
Dam safety evaluation with the SANCOLD freeboard, the register of dams and the compliance position of your own.
- evaluate_dam_safety
- search_registered_dams
- get_dam_compliance
Projects and review
File every run in a project, follow a tracked workflow step by step, and record the engineer’s ruling on a flood study.
- create_project
- get_project
- start_workflow
- run_workflow_step
- record_flood_review
The full tool list is in the guide, with the ready-made procedures the server publishes for a peak flow from scratch, sizing a culvert and reviewing a flood study.
On its own, ChatGPT guesses. Connected to Studio, it calculates.
Design software has always been worked by hand: one form, one value at a time. Agents change who does that work. An AI model can talk about hydraulics, but it cannot be trusted with the arithmetic, and it leaves nothing behind for a reviewer. An agent needs engines it can call, records it can leave and limits it cannot argue its way past. Studio is built for that.
ChatGPT or Claude on its own, asked to design a culvert
- The number comes out of a language model. It can be wrong and still read as right.
- No method is named and no inputs are kept, so there is nothing to check it against.
- Ask the same question twice and you can get two answers.
- The answer stays in a chat window. You retype it into your design.
The same agent, connected to Delta Hydro Studio
- The agent calls the engine: FHWA HDS-5 for the culvert, the Rational method for the peak. The model supplies inputs and reads results.
- Every input and result is saved as a record in your project.
- The same inputs give the same result, every time.
- The record opens in Studio with charts, intermediate tables and a Word and Excel report.
The data entry goes away
A flood study is a long chain of forms: delineate, read off the rainfall, fill in each method, size the structure, copy it all into a report. An agent works through the same steps in the same order, and stops at each decision that is yours.
Two hundred sites, one instruction
Hand the agent a list of crossings. One submission runs a flood study for up to 200 outlets, each with the methods its own catchment allows, and one table compares the sites. Flood studies cover South African sites.
The engineer becomes the reviewer
Your hours go to checking catchments, questioning assumptions and choosing the design flood. The agent recommends. A ruling is recorded only when you state it, under your name.
Brief it like a colleague
Give the agent what you would give a person: where, which return period, and the levels and limits you know. It asks for an engineering value it does not have; it is told never to make one up.
“Delineate the catchment at this outlet and give me the 1:100 peak by the Rational method and by TR-55.”
“Size a box culvert for 21 m³/s at this crossing. Road level 103.20 m, HW/D no more than 1.2.”
“Here are 40 crossings on the route. Quote a flood study for each and submit them when I approve the total.”
“Compare the methods in last week’s flood study and tell me which you would set aside, and why.”
The engineer stays in charge
An agent that can spend credits and save records needs limits that do not depend on its good behaviour. These are enforced by the server.
A quote before every paid study
A flood study takes two calls. The agent gets the price, puts it to you, and submits only once you approve it.
A spending limit on every key
Each API key has an hourly limit on what agent calls may spend: 30 credits unless you set another. Past it, calls are refused and nothing is charged.
A brake on runaway loops
After eight saved records from one tool in ten minutes, the server stops the agent and makes it ask you before it carries on.
Nothing is deleted
Agent tools read your data or add new records. None of them can delete a record or a project.
Agent work is marked
Every record an agent creates carries its origin in the calculation audit trail, so a reviewer can see which results came from an agent.
You are the engineer of record
An agent cannot take professional responsibility. Review every input it chose and every result before any of it reaches a drawing or a report.
Questions about agents and MCP
What is an MCP server?
Which AI agents can connect?
Does Studio have its own AI assistant?
Can my agent use data from QGIS or my own files?
How is this different from asking ChatGPT or Claude to do the calculation?
What does it cost?
Can the agent change or delete my work?
Who is responsible for the result?
Is it available now?
Give your agent real engines
You need a Studio account and an AI client that supports MCP. The guide takes you from there to the first saved record.