Using OGR Slip2D with an AI agent

Edit claude_desktop_config.json, which is in %APPDATA%\Claude\ on Windows and in ~/Library/Application Support/Claude/ on macOS; the Microsoft Store version keeps it in %LOCALAPPDATA%\Packages\Claude_<id>\LocalCache\Roaming\Claude\. Restart Claude afterwards.

{
  "mcpServers": {
    "ogr-slip2d": {
      "command": "ogr-slip2d-mcp",
      "args": ["--workdir", "C:\\path\\to\\projects"]
    }
  }
}

If Claude cannot find the command, put the full path of ogr-slip2d-mcp.exe (inside your virtual environment's Scripts folder) in command.

One command registers the server for Claude Code:

claude mcp add ogr-slip2d -- ogr-slip2d-mcp --workdir "C:\path\to\projects"

Claude Code asks for confirmation before each tool call. Leave that confirmation on for python_exec (see Security).

LM Studio reads the MCP servers for its local models from %USERPROFILE%\.lmstudio\mcp.json on Windows, or from ~/.lmstudio/mcp.json on macOS and Linux:

{
  "mcpServers": {
    "ogr-slip2d": {
      "command": "ogr-slip2d-mcp",
      "args": ["--profile", "compact", "--workdir", "C:\\path\\to\\projects"]
    }
  }
}

With small models, use --profile compact. It publishes 14 tools instead of 74: model_define builds the geometry in one call and python_exec covers the rest. A small model gets lost among many tools, and each tool description takes up context in every request.

Ollama serves models but does not speak MCP itself; that is the job of the client that uses the model. With mcp-client-for-ollama, or any MCP client that accepts the mcpServers format, the entry is the same as for LM Studio:

{
  "mcpServers": {
    "ogr-slip2d": {
      "command": "ogr-slip2d-mcp",
      "args": ["--profile", "compact", "--workdir", "C:\\path\\to\\projects"]
    }
  }
}

Open WebUI connects to MCP servers over HTTP only (Streamable HTTP, Open WebUI 0.6.31 or newer). Start the server with a token:

ogr-slip2d-mcp --transport http --port 8765 --token-file token.txt --workdir "C:\path\to\projects"

Then, in Open WebUI, go to Settings → Admin → Integrations → External Tool Servers → + Add Connection and set the type to MCP (Streamable HTTP), the URL to http://127.0.0.1:8765/mcp and the authentication to Bearer, with the token as the key.

Without --token-file or OGR_MCP_TOKEN, the server generates a token and prints it when it starts. If Open WebUI runs in Docker, use http://host.docker.internal:8765/mcp and start the server with --host 0.0.0.0 and an explicit token; with that host and no token, the server refuses to start.

1. Start the server on loopback with a token of at least 32 characters and the public address of your tunnel:

ogr-slip2d-mcp --transport http --port 8765 --token-file token.txt --public-url https://YOUR-TUNNEL.example --workdir "C:\path\to\projects"

2. Put a tunnel that terminates HTTPS in front of http://127.0.0.1:8765. For example, cloudflared tunnel --url http://127.0.0.1:8765 gives you the address for --public-url (restart the server with it).

3. In ChatGPT, enable developer mode and create a connector with the address https://YOUR-TUNNEL.example/mcp and OAuth authentication.

4. ChatGPT sends you to the server's /approve page, where you check that the client is the one you expect and type the token from token.txt. The server acts as its own OAuth 2.1 authorisation server (PKCE, dynamic registration). Access tokens last one hour and are kept only in memory.

With the OpenAI Agents SDK for Python, the server is used as a stdio MCP server:

from agents import Agent, Runner
from agents.mcp import MCPServerStdio

async with MCPServerStdio(
    params={"command": "ogr-slip2d-mcp",
            "args": ["--workdir", r"C:\path\to\projects"]},
    client_session_timeout_seconds=60,   # an analysis can take a while
) as ogr:
    agent = Agent(name="geotechnics", mcp_servers=[ogr],
                  instructions="Analyse slopes with OGR Slip2D.")
    result = await Runner.run(agent, "Build and analyse the slope ...")

Set client_session_timeout_seconds explicitly. The SDK's default is a few seconds, while analysis_run waits up to 30 s by default before returning a job to poll with job_get.

Clients that integrate tools through OpenAPI (for example BionicGPT) can use mcpo, which turns an MCP server into an OpenAPI service. Install it in a separate environment, pinned to the 1.x MCP SDK it was built for:

python -m venv mcpo-env
mcpo-env\Scripts\pip install mcpo "mcp<2"
mcpo-env\Scripts\mcpo --port 8000 --api-key "your-key" -- ogr-slip2d-mcp --workdir "C:\path\to\projects"

The specification is then served at http://localhost:8000/openapi.json, with interactive documentation at /docs. model_render returns a data URI; pass save_path if you also want the PNG on disk.