openLCA MCP Server

An MCP (Model Context Protocol) server that connects AI assistants to a running openLCA instance. Developed and tested with Claude Desktop; compatible with any MCP client that supports stdio transport. Built by Below280, the UK partner for openLCA.

The server exposes 31 tools covering the full LCA workflow: exploring databases, building and editing models, running calculations (scenarios, sensitivity, Monte Carlo, contribution analysis), auditing and validating models, and extracting data quality assessments. All calculation patterns are tested against production ecoinvent databases.

The server works with both ecoinvent-family databases (ecoinvent, EN15804GD, HiQLCD, BAFU) and FLCAC-family databases (LCA Commons, US LCI, USEEIO). It asks which family you are using, or auto-detects from the flow property names.

Available on PyPI and the MCP Registry.

Install

Quick setup (PyPI)

pip install b280-olca-mcp

This installs the server and its dependencies (mcp, olca-ipc) in one step. The server runs from anywhere with python -m b280_olca_mcp.

Development setup (GitHub)

Clone the repository for the latest code, the React dashboard, or if you want to modify the server:

git clone https://github.com/Below280/B280-olca-MCP.git
cd B280-olca-MCP
pip install -r requirements.txt

The server entry point is b280_olca_mcp/server.py.

What it does

Someone using this MCP can say things like:

• 'Make me a model with 34 kWh UK electricity, 5 kg sodium hydroxide, and 34 kWh steam'
• 'Build an EPD model from this LCI'
• 'Run two scenarios, one with transport at 100 km and one at 500 km'
• 'Which processes contribute most to climate change in my system?'
• 'Vary the electricity and PET resin parameters by 10%'
• 'Validate my product system and check if the linking is correct'
• 'Run 1000 Monte Carlo iterations and show me the uncertainty'
• 'Help me connect to openLCA from R / Go / Rust'
• 'What can you do?' (the help tool)

The AI assistant handles the conversation, builds the tool calls, and presents results visually with charts, tables, and exportable data.

Connect your MCP client

The server uses stdio transport. Any MCP client that can spawn a local Python process will work.

Claude Desktop

In Claude Desktop, go to Settings > Developer > Edit Config. Add an openLCA entry to the mcpServers section.

If you installed via PyPI:

{
  "mcpServers": {
    "openLCA": {
      "command": "python",
      "args": ["-m", "b280_olca_mcp"],
      "env": {
        "OLCA_PORT": "8080"
      }
    }
  }
}

If you cloned from GitHub:

{
  "mcpServers": {
    "openLCA": {
      "command": "python",
      "args": ["C:/path/to/B280-olca-MCP/b280_olca_mcp/server.py"],
      "env": {
        "OLCA_PORT": "8080"
      }
    }
  }
}

On Windows, if Claude Desktop cannot find Python, use the full path to your Python executable.

Restart Claude Desktop after saving.

Cursor

Go to Settings > Tools & MCP > Add MCP Server. Choose stdio transport, set the command to python and the argument to -m b280_olca_mcp (PyPI) or the path to b280_olca_mcp/server.py (GitHub clone). Cursor picks up config changes without restarting.

VS Code

Add the server to .vscode/mcp.json in your workspace or user settings:

{
  "servers": {
    "openLCA": {
      "type": "stdio",
      "command": "python",
      "args": ["-m", "b280_olca_mcp"],
      "env": {
        "OLCA_PORT": "8080"
      }
    }
  }
}

Note: VS Code uses servers as the root key, not mcpServers. If using GitHub Copilot, switch Copilot Chat to Agent mode.

ChatGPT

ChatGPT cannot directly launch a local stdio server. OpenAI provides the Secure MCP Tunnel, which bridges ChatGPT to a local MCP server without exposing anything publicly. The architecture is:

ChatGPT → Secure MCP Tunnel → B280 MCP (stdio) → openLCA IPC localhost:8080

The tunnel launches the B280 server (python -m b280_olca_mcp) on your machine. No changes to the B280 server are needed; it remains stdio throughout. No additional search or fetch tools are required for ChatGPT custom MCP servers.

B280's existing tool annotations (read-only, destructive, idempotent) are used by ChatGPT when deciding whether to auto-approve or prompt for confirmation.

For the full setup guide, including installation, configuration, troubleshooting and security considerations, see docs/chatgpt.md.

ChatGPT support is currently experimental. A remote Streamable HTTP transport option may be added in future for users who prefer not to run the local tunnel.

Other MCP clients

Any client that spawns a local Python process over stdin/stdout should work. With PyPI: python -m b280_olca_mcp. With GitHub clone: python path/to/b280_olca_mcp/server.py.

Start it

1. Open your database in openLCA
2. Start the IPC server (Tools > Developer Tools > IPC Server > green play button, port 8080)
3. Restart your MCP client (or reconnect)
4. Ask something like 'what can you do?' or 'what's in my openLCA database?'

After creating or modifying anything, press the Refresh button in openLCA's toolbar to see changes in the GUI.

Data security

The MCP server runs locally and communicates with openLCA on localhost. The AI client (Claude Desktop or equivalent) sends tool results to its provider's servers for processing. This means process names, exchange data, parameter values, and impact results from your database will be in the conversation.

Do not connect a confidential client database through a personal or free-tier AI account. Use a business or enterprise account with appropriate data retention controls, and check your provider's data processing terms before connecting any database containing sensitive information.

Tools

Explore (11 tools)

Build (6 tools)

Audit (5 tools)

Calculate (8 tools)

Meta (1 tool)

Tool annotations

Every tool carries MCP tool annotations (readOnlyHint, destructiveHint, idempotentHint, openWorldHint) so clients can decide whether to auto-approve or prompt for confirmation. Explore and Calculate tools are read-only. Build tools signal that they modify the database. delete_entity is marked destructive. CSV tools are marked as writing files to disk. chemical_synonyms is marked as reaching an external service (PubChem).

Resources

The server exposes four MCP resources that AI clients can read on demand:

The IPC protocol reference enables AI clients to help users connect to openLCA from any programming language by generating client code from the protocol specification.

EPD model building

The server includes a guided workflow for building EN15804 EPD models from LCI data. The AI maps LCI items to lifecycle modules (A1-D), creates the folder structure, bridge processes, module processes, and product systems. Ask 'build an EPD from this LCI' to start.

Standalone scripts

For repeated analyses, the server recommends standalone Python scripts that run without AI tokens. After a scenario or sensitivity calculation, it offers to generate the CSV file needed to run the equivalent script from the openLCA-IPC-tools-python repository.

Related repositories

About

Built by Below280 Limited, a UK LCA, EPD and CBAM consultancy and official UK partner for openLCA. The calculation patterns in this server are derived from production scripts used in EPD and LCA consulting work, tested against ecoinvent 3.10/3.11/3.12, EN15804GD, and US Federal LCA Commons (FLCAC) databases.

For issues: GitHub Issues or mcp-feedback@below280.com

Documentation

• Connecting AI to openLCA (full setup guide)
• openLCA scripting knowledge base
• openLCA Manual

Licence

MPL-2.0

mcp-name: io.github.Below280/b280-olca-mcp