Use fortranspire from Claude Code

fortranspire ships a ready-to-install Claude Code skill so Claude can analyze, port, and document Fortran 90 code from inside the IDE — no MCP plumbing required.

Local skill (recommended)

The skill is skills/claude-code/fortranspire/ in the repo. Copy it into your project’s .claude/skills/:

pip install fortranspire           # the actual binary
# In whatever Claude-Code project you want to use it from:
mkdir -p .claude/skills
cp -r <fortranspire-repo>/skills/claude-code/fortranspire .claude/skills/

…then reload Claude Code. The /fortranspire slash command appears, and Claude auto-invokes the skill on Fortran-related requests.

The skill’s SKILL.md documents every subcommand and includes a decision flow — Claude always runs fortranspire explain first (zero LLM cost, tells you whether the file is portable) before recommending the paid gpu or translate verbs.

What needs the LLM, and what doesn’t

Verb	LLM call?	API key needed?
analyze	No	No
explain	No	No
format	No	No
graph	No	No
diff	No	No
report	No	No
bench	No	No
doc --no-llm	No	No
doc --with-llm	Yes	Yes
gpu	Yes	Yes
port-batch	Yes	Yes
translate	Yes	Yes

For LLM verbs, see LLM endpoints.

Remote MCP (advanced)

If your Claude Code is configured to talk to an MCP server (e.g., on a shared GPU host), point it at a running fortranspire mcp:

# On the server
fortranspire mcp --port 8000
# or with env vars for the container
MCP_HOST=0.0.0.0 MCP_PORT=8000 fortranspire mcp

…and register http://<host>:8000/sse in Claude Code’s MCP settings. The server exposes nine tools — the same surface as the CLI minus the format / diff / report verbs which are best run locally on the user’s filesystem:

• analyze_kernels — static analysis, optional SARIF output

• explain_port_cost — pre-flight estimate (no LLM)

• build_call_graph — Mermaid call-graph

• generate_docs — !> docstring injection (no-LLM or LLM-driven)

• translate_kernel_gpu — Phase 1 port

• translate_kernel — Phase 2 JAX translation

• profile_kernels — performance comparison

• ask_agent — natural-language query

• agent_status — server config dump

When FORTRANSPIRE_TOKENS_FILE or API_KEY is set, the MCP server runs authenticated with rate-limiting + HMAC-signed audit logs (docs/concepts/security.md).

Troubleshooting

• /fortranspire not in the slash-command menu — verify .claude/skills/fortranspire/SKILL.md exists and Claude Code has been reloaded (close and re-open the project).

• Claude says “fortranspire: command not found” — pip install fortranspire in the same venv Claude Code uses, or in ~/.local/bin if you’re using system Python.

• LLM verb fails with “no MISTRAL_API_KEY” — add the key to .env at the project root, or export it in the shell Claude Code spawns (~/.zshrc, ~/.bashrc).