Use fortranspire from mistral-vibe¶

mistral-vibe — Mistral’s CLI coding agent — speaks the Model Context Protocol natively. fortranspire ships a FastMCP server with both stdio (vibe spawns it per session — zero ports, zero auth) and HTTP/SSE (a permanent service for CI / LAN / Claude Desktop) transports.

The recommended path on a developer Mac (or a Mac mini acting as your self-hosted runner) is stdio: vibe boots the server when you cd into a trusted project, the tools appear in your toolbox, and they vanish when you exit. No daemon to babysit.

This is the strongest demonstration of the sovereignty story (see Architecture vs LLM): Loki + the deterministic harness absorb 60-70% of the work, and the LLM (Mistral on mistral-vibe, Claude on Claude Code) is a fungible component.

Prerequisites¶

brew install mistral-vibe                       # installs `vibe` at /opt/homebrew/opt/mistral-vibe/bin/vibe
pip install fortranspire                        # installs `fortranspire` console script
fortranspire mcp --stdio < /dev/null            # sanity-check: returns immediately, no banner

The third command should exit cleanly with no output — FastMCP is waiting for JSON-RPC on stdin and shuts down on EOF.

Self-hosted runner on a Mac mini (stdio)¶

This is the path you want if vibe and the MCP server live on the same host. Edit ~/.vibe/config.toml and append:

[[mcp_servers]]
name      = "fortranspire"
transport = "stdio"
command   = "fortranspire"
args      = ["mcp", "--stdio"]

# Optional — the server reads these to call out to Mistral when you
# invoke an LLM verb (translate_kernel_gpu, generate_docs --with-llm).
[mcp_servers.env]
MISTRAL_API_KEY = "${MISTRAL_API_KEY}"
MISTRAL_MODEL   = "codestral-latest"

Path resolution. If which fortranspire is empty in vibe’s environment (Homebrew shells out without your full profile), pin the absolute path instead — e.g. command = "/Users/you/.venvs/fortranspire/bin/fortranspire".

Trust the project directory so vibe will let the MCP server see your sources:

vibe --trust --workdir ~/work/phyex            # one-shot
# or persist:
# vibe                                          # then `/trust` inside the TUI

Verify the registration from inside a vibe session:

/mcp                                           # lists discovered servers
/tools                                         # the 9 fortranspire_* tools should appear

Talking to it in natural language¶

Once the server is registered, vibe’s planner can call the tools from plain French/English. A typical “is this kernel portable?” session:

> Regarde phyex/src/MNH/rain_ice.f90 et estime le coût d'un portage GPU.

vibe will call fortranspire_explain_port_cost, then summarise the table it gets back — routine count, control-flow complexity, GPU portability score, expected token spend for a Phase-1 port. No source file leaves the process; only the rendered report flows to the model.

A few prompts that map cleanly onto the exposed tools:

Intent	Tool that fires
“Audite ce fichier” / “any GPU-unfriendly patterns?”	`analyze_kernels`
“Combien coûterait un port ?” / “is it worth porting?”	`explain_port_cost`
“Dessine le graphe d’appels”	`build_call_graph`
“Documente ces routines”	`generate_docs`
“Génère le wrapper OpenACC + Cython”	`translate_kernel_gpu`
“Prototype une version JAX”	`translate_kernel`

The full surface and arguments:

Tool	What it does	LLM call?
`analyze_kernels`	Loki AST analysis, optional SARIF output	No
`explain_port_cost`	Pre-flight cost & risk estimate	No
`build_call_graph`	Mermaid call-graph	No
`generate_docs`	`!>` docstring injection (no-LLM or LLM-driven)	Optional
`translate_kernel_gpu`	Phase 1 — Fortran → OpenACC + Cython	Yes
`translate_kernel`	Phase 2 — Fortran → JAX (experimental)	Yes
`profile_kernels`	Performance benchmarking	No
`ask_agent`	Natural-language query against the code	Yes
`agent_status`	Dump server config	No

5-minute demo on PHYEX¶

PHYEX bundles the Météo-France physics schemes (microphysics, turbulence, convection). The kernels are small, self-contained, and ideal for a guided demo.

git clone https://github.com/UMR-CNRM/PHYEX ~/work/phyex
cd ~/work/phyex
vibe --trust .

Then, in the vibe session, a guided three-step demo:

1. Triage — is this kernel portable?

> Compare ces deux fichiers pour un portage GPU :
>   src/common/turb/mode_compute_function_thermo.F90
>   src/common/turb/mode_tke_eps_sources.F90
> Lequel est portable tel quel ? Lequel demande un refactor d'abord ?

vibe calls fortranspire_explain_port_cost sur chaque fichier. La sortie surface FORT001 — I/O in kernel candidate sur le second (un PRINT dans TKE_EPS_SOURCES qui bloque le port GPU), et un coût propre (~$0.01) sur le premier.

2. Visualiser la structure du portable

> Dessine le call-graph de mode_compute_function_thermo.F90.

fortranspire_build_call_graph rend un Mermaid flowchart LR embarquable directement dans une PR ou un mdBook.

3. Générer le wrapper OpenACC + Cython

> Génère le port OpenACC + Cython pour mode_compute_function_thermo.F90.

fortranspire_translate_kernel_gpu enchaîne le pipeline Phase-1 et écrit output/fortran_gpu/*.f90 + output/cython/*.pyx à côté du cwd. vibe propose ensuite la compilation + validation via gfortran.

Tip. Toujours commencer par explain_port_cost avant d’appeler un verbe LLM — ça évite de griller des tokens sur un fichier qui se fera retoquer en validation pour une raison structurelle (I/O, COMMON, pointeurs).

Alternative — permanent HTTP service (CI / LAN access)¶

When you want the same server reachable from Claude Code Desktop, from your laptop hitting the Mac mini over the LAN, or from a CI runner:

fortranspire mcp                                # default: SSE on $MCP_HOST:$MCP_PORT
# or
MCP_HOST=0.0.0.0 MCP_PORT=8000 fortranspire mcp

The server listens on SSE at http://<host>:8000/sse. Auth is opt-in:

export FORTRANSPIRE_TOKENS_FILE=/etc/fortranspire/tokens.json
export FORTRANSPIRE_AUDIT_SECRET=<random-256-bit-secret>
fortranspire mcp

See security for the token registry format and the HMAC-signed audit log.

vibe’s HTTP registration looks like:

[[mcp_servers]]
name      = "fortranspire"
transport = "streamable-http"
url       = "http://mac-mini.local:8000/sse"

[mcp_servers.auth]
type        = "static"
api_key_env = "FORTRANSPIRE_TOKEN"

To keep the server alive across reboots, drop a LaunchAgent under ~/Library/LaunchAgents/fr.fortranspire.mcp.plist:

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
  <key>Label</key><string>fr.fortranspire.mcp</string>
  <key>ProgramArguments</key>
  <array>
    <string>/usr/local/bin/fortranspire</string>
    <string>mcp</string>
  </array>
  <key>EnvironmentVariables</key>
  <dict>
    <key>MCP_HOST</key><string>0.0.0.0</string>
    <key>MCP_PORT</key><string>8000</string>
    <key>MISTRAL_API_KEY</key><string>...</string>
  </dict>
  <key>RunAtLoad</key><true/>
  <key>KeepAlive</key><true/>
  <key>StandardOutPath</key><string>/tmp/fortranspire-mcp.out.log</string>
  <key>StandardErrorPath</key><string>/tmp/fortranspire-mcp.err.log</string>
</dict>
</plist>

Then launchctl load -w ~/Library/LaunchAgents/fr.fortranspire.mcp.plist.

Why both Claude Code and mistral-vibe?¶

Surface	LLM	Sovereignty
Claude Code	Claude (US)	Mixed — code stays local, model is US-hosted
mistral-vibe	Mistral (EU)	Full — code + inference both EU-resident
`fortranspire mcp` self-hosted	Mistral self-hosted	Air-gapped possible

Same agent, three sovereignty postures. Choose per project / per customer requirement.

Known gaps¶

The MCP server returns plain-text responses (Markdown-friendly). Structured-output support (Pydantic JSON schemas) is on the roadmap — track progress in the issue tracker.
The translate_kernel_gpu tool runs the full Phase-1 LangGraph in the same process as the MCP server. On large kernels (>2k LoC) this can hold vibe’s UI for a few minutes — prefer running the CLI fortranspire gpu file.f90 for batch work.