Configuration

fortranspire reads its configuration exclusively from environment variables (or from a local .env file, loaded automatically). A complete template is shipped at the root of the repository as .env.example.

LLM endpoint

Variable	Default	Description
MISTRAL_ENDPOINT	https://api.mistral.ai/v1	Base URL of any OpenAI-compatible chat endpoint
MISTRAL_API_KEY	(unset)	Bearer token sent to the endpoint
MISTRAL_MODEL	mistral-large-latest	Model name passed in every request
LLM_TEMPERATURE	0.0	Sampling temperature
LLM_TOP_P	0.9	Nucleus sampling cutoff
LLM_NUM_PREDICT	2048	Maximum tokens per completion

The same variables work for Mistral La Plateforme, a self-hosted vLLM server, an Ollama instance, or any other OpenAI-compatible backend.

MCP server

Variable	Default	Description
MCP_HOST	0.0.0.0	Bind address
MCP_PORT	8000	Listen port
API_KEY	(unset)	If set, require Authorization: Bearer <key> on every
		request (skipped on /health and /ping)

Pipeline behavior

Variable	Default	Description
WORKSPACE_DIR	repo root	Where intermediate files are written
MAX_ITERATIONS	10	Maximum LangGraph re-entry count
OUTPUT_DIR	output/	Root for generated Fortran / Cython / JAX

Compiler toolchain

Variable	Default	Description
FC_CPU	gfortran	Compiler used for the CPU syntax checks
FC_GPU	nvfortran	Compiler used for the OpenACC validation step
CUDA_ARCH	cc80	GPU architecture passed to nvfortran -gpu=

Loading the .env

The CLI and the MCP server both call python-dotenv at startup, so simply placing a .env file at the repository root (or in your working directory) is enough — no flag required. CI workflows should set the same variables through repository secrets.