pyvar-mcp

MCP server exposing all 385 pyvar.com risk functions — VaR, credit risk, derivatives, liquidity, operational risk, portfolio analytics, ALM, regulatory capital — as Claude Code tools.

It's a thin wrapper over the live pyvar API (each tool call is a real HTTPS request), not a bundled copy of the compute engine — results always match what's actually deployed at pyvar.com.

Install

/plugin marketplace add fibtecltd/pyvar
/plugin install pyvar-mcp@pyvar-marketplace

You'll be prompted for a pyvar API key during install. Sign in at https://www.pyvar.com#get-api-key (emailed link or Google; no password, no credit card) and create one on the dashboard. It starts with pyvar_, lasts a year, and you can revoke it there at any time.

The plugin starts the server with uvx pyvar-mcp, which downloads it from PyPI on first use, so it needs uv on your PATH. If the tool calls fail with a "command not found" style error, a missing uv is almost certainly why. No uv? Run pip install pyvar-mcp and change the plugin's mcpServers command to pyvar-mcp (with no args). Working from a checkout of this repo instead: pip install -e plugins/mcp.

pyvar-mcp is published to PyPI by .github/workflows/pyvar-mcp-publish.yml when a pyvar-mcp-v* tag is pushed (trusted publishing; the tag must equal the version in pyproject.toml).

Tools

Two generic tools, meant as your first choice:

• list_pyvar_functions(domain?) — browse available functions, optionally filtered to one domain (alm, credit-risk, derivatives, liquidity, market-risk, operational, portfolio, regulatory).
• call_pyvar_function(domain, function_name, params) — call any function by name. Validates params against the function's own schema before calling the API.

Plus all 385 functions as individually named tools (alm_stress_test, historical_simulation_var, etc.) with precise per-parameter schemas, for when you already know exactly which one you want.

Local development

cd plugins/mcp
pip install -e ".[dev]"
pytest

Regenerating the tool catalogue after portal/functions.json changes:

python3 scripts/generate_mcp_tools.py   # from the repo root

.github/workflows/plugins-ci.yml fails CI if the committed pyvar_mcp/_generated/ doesn't match what regenerating produces — same drift-check pattern the skills plugins and pyvar-client use. The catalogue is one module per domain, not one file: the plugin directory's safety scan reads a file in full only up to 256 KiB, and the generator refuses to write a file over 200,000 bytes (the largest today is about 78,000).