SysML v2 Co-Pilot

sysml — SysML v2 Systems-Engineering Co-Pilot (Claude Code plugin)

An AI-native SysML v2 co-pilot. It authors valid textual SysML v2 (.sysml), grounds every type/unit in the standard library (ISQ, SI, …) instead of inventing them, and validates with the real SysML v2 Pilot compiler in a self-correcting loop so the output actually compiles.

What's inside

sysml/
├── .claude-plugin/plugin.json     # plugin manifest (auto-loads as `sysml@skills-dir`)
├── .mcp.json                      # declares the `sysml` MCP server
├── skills/sysml-copilot/          # the co-pilot skill (SKILL.md + references + examples)
└── mcp-server/                    # the MCP server (validate + library introspection)
    ├── server.py                  # stdio MCP server (run via `python3`)
    ├── library_index.py           # name→element index over the standard library
    ├── java/SysmlValidatorServer.java   # warm Pilot-kernel validator (stdin/stdout JSON)
    └── setup.sh                   # downloads the kernel jar + compiles the validator

MCP tools

• validate_sysml_file(content | path, context_paths?) → {ok, diagnostics:[{line, column, severity, code, syntax, message}]}
• query_library(query, limit?) → ranked standard-library elements with qualified names
• get_library_element(qualified_name) → an element's declaration excerpt

Prerequisites & setup

• Java 21+ and a python3 (>=3.10) on PATH. The MCP server is stdlib-only — no Python dependencies to install, no venv, no uv.
• The SysML v2 standard library (clone Systems-Modeling/SysML-v2-Release).
• The Pilot kernel fat jar (downloaded automatically by setup.sh).

# one-time: fetch the kernel jar and compile the validator
mcp-server/setup.sh

Configure paths via env if you need to override the defaults (which land in plugin-local .runtime/):

SYSML_LIBRARY_PATH=${CLAUDE_PLUGIN_ROOT}/mcp-server/.runtime/sysml.library
SYSML_KERNEL_JAR=${CLAUDE_PLUGIN_ROOT}/mcp-server/.runtime/jupyter-sysml-kernel-0.59.0-all.jar

Launch configuration (.mcp.json)

The MCP server is stdlib-only and launched directly by a system python3 (no venv, no uv — those caused handshake timeouts -32000 when deps were resolved at launch). The .mcp.json uses ${CLAUDE_PLUGIN_ROOT}-relative paths so the plugin is portable and works immediately after running setup.sh.

Run mcp-server/setup.sh once (downloads the kernel jar if missing, compiles the Java validator), launch Claude Code from the repo root, then /reload-plugins and verify /plugin shows sysml connected. The Java validator also auto-compiles on first use if setup.sh wasn't run.

Notes

• The validator runs as a warm process: the JVM + 94-file library load happens once, then validations run in well under a second. Cold-starting per call would be unusable in a loop.
• The full standard library and the kernel jar are referenced by path, not vendored — they are large and carry the upstream copyright-holders' license. Only small grounding assets (curated syntax reference, BNF grammar, example models) are bundled in the skill.