An AI copilot for calculations on geological models, alongside Petrel and RMS. Open an exported model, calculate scenarios and check the results in plain words from any MCP client. It works today in Claude Desktop and Claude Code. Models and computation stay on your workstation.
Code computes the numbers, not the language model.
You confirm the assumptions.
You build the model where you always have. Kanysh Basic takes the exported files — DATA/GRDECL, surfaces, wells, LAS — and does the work that usually goes to Excel and manual checks: regional volumes, scenarios, reconciliation and QC. There is one control: conversation.
Five core workflows are ready MCP prompts, plus large models and running your own deck in OPM Flow. No tool names to learn: describe the task in your own words, and the agent picks the tool and asks for what is missing.
“Check the model and show issues with severity and IJK”
“Calculate oil and free gas by EQLNUM with my contacts”
“Compare the volumes with my Petrel CSV and explain the differences”
“Run a porosity tornado for two regions”
“Compare two model versions and the pore-volume impact”
“Export the report to Excel and PDF and show the model in 3D”
“Import the EGRID and INIT and check the model”
“Run my deck in OPM Flow and compare with ECLIPSE”
The LLM in your MCP client understands the task and calls Kanysh Basic’s typed tools. The numbers are computed by tested Python code on your machine’s CPU, and every result is checked by an independent method.
folder_assess identifies file roles and proposes contacts, PVT and units — with file and line.Kanysh Basic is a local MCP server with an offline viewer. It has no built-in LLM and no cloud. Your MCP client decides which messages reach its AI provider.
The public Norne field model (Statoil / OPM): 113,344 cells, 44,927 active. The whole path, from folder assessment to PDF, ran through a real MCP session. Fluid parameters are explicit demo assumptions, not a reserves estimate.
| EQLNUM | Cells | OWC, m | STOIIP, sm³ | GIIP, million sm³ |
|---|---|---|---|---|
| 1 | 4 224 | 2692 | 9 347 774 | 5 375 |
| 3 | 1 296 | 2618 | 2 860 856 | 822 |
| Total | 5 520 | — | 12 208 630 | 6 197 |
Model QC: ISSUES — 243 degenerate cells and 1 twisted cell. They are not silently removed: volumes are calculated for QC-compatible regions 1 and 3. Demo assumptions: Bo 1.2, Bg 0.005, Sw 0.25. Norne data: Statoil / OPM, ODbL-1.0 / DbCL-1.0.
We test Kanysh Basic beyond the demo. A synthetic corner-point model with 200 million cells runs the whole path on an ordinary laptop, and the real Equinor Volve model showed where we match ECLIPSE and where gaps remain.
| Stage | Time | Peak RAM |
|---|---|---|
| Import DATA/COORD/ZCORN | 17.7 min | 12,1 GiB |
| Model QC | 12.3 min | 2,6 GiB |
| Regional volumes | 14.6 min | 2,1 GiB |
| Tornado | 13.8 min | 2,1 GiB |
Laptop: 16 threads, 64 GB RAM, Windows 11, CPU. A single run; all checks PASS, with three deliberately broken cells excluded explicitly. Method and raw data.
Equinor Volve data (Equinor Open Data Licence) is not redistributed with Kanysh. The deck needed 3 documented edits for OPM to accept it; OPM ignored ADDZCORN and PRIORITY. Full report.
For Claude Desktop it is a one-click .mcpb extension. For other MCP clients the same archive runs as a regular stdio server. uv installs the Python dependencies; the archive holds the code and the public Norne demo, nothing else.
Clone the repository, install dependencies with uv (Python 3.12 required) and register a stdio server in your client’s configuration (on Windows the path is .venv\Scripts\kanysh-basic.exe). No git? The .mcpb file is a regular zip: unpack it and run uv run --directory <folder> server.py.
ChatGPT on the web and in its app connects only to remote MCP servers over HTTPS, so a local server is out of reach. Connecting via OpenAI Secure MCP Tunnel is planned.
A numerical PASS does not replace geological judgement. Here is where the limits are today.
Kanysh Advanced is the next step: an AI-native workstation in the browser, on the customer’s GPU server, on top of OSDU. Seismic, wells and modelling through a dialogue with the agent. This is a development direction, not a delivery promise: Basic does not include these features.
MCP copilot alongside Petrel/RMS: import, QC, volumes, reconciliation, sensitivity, reports.
Multi-horizon, multi-fault structural models, well-log upscaling, facies and spatial property simulation, well correlation.
Large ROFF, remote GPU compute, S3 and OSDU, server-side rendering of large scenes, seismic interpretation.
Basic already runs your finished deck in OPM Flow. Next: building the flow model, history matching (ES-MDA), ensembles and development-plan optimisation.
Basic shows the working core: verifiable calculations from a conversation. We are open to pilots on real models and to conversations with investors and strategic partners, especially those who know Petrel/RMS workflows, OSDU and oil and gas interpretation.
In short: what it is, where the data lives, what you need to run it.
simulation_deck_run tool runs your finished ECLIPSE deck in OPM Flow exactly as supplied: nothing is generated or defaulted. It needs Linux, or Windows with WSL2 — kanysh-basic setup-opm installs OPM into a dedicated distribution and self-tests it on SPE1. Without a set-up OPM the request refuses honestly, with install steps. Building the flow model and history matching belong to Kanysh Advanced.