AI Coding Tools (MCP)¶

MEMO ships a Model Context Protocol server. Any MCP client — Cursor, Claude Code, Windsurf — can use it to query your model while it edits your .sysml files.

This matters because a coding agent editing SysML by hand is working blind. It cannot see which kinds your ontology defines, which relationship types are legal, or what the validator will say about its change. With the MCP server it can look all of that up, and check its own work afterwards.

Setup¶

Cursor¶

From your project directory:

memo mcp init

That writes two files:

File	Purpose
`.cursor/mcp.json`	Registers the `memo` server with Cursor
`.cursor/rules/memo.mdc`	MEMO conventions, applied when Cursor touches a `.sysml` file

Restart Cursor, then check Settings → MCP for a server named memo.

If you already have MCP servers configured, memo mcp init merges into the existing .cursor/mcp.json rather than replacing it — your other servers are left alone.

To skip the rules file:

memo mcp init --no-rules

Claude Code¶

claude mcp add memo -- npx -y --package=@memoarchitect/tools memo mcp --project /path/to/your/project

memo mcp init prints this line with your project path already filled in.

Any other MCP client¶

The server speaks MCP over stdio. Point your client at:

npx -y --package=@memoarchitect/tools memo mcp --project /path/to/your/project

No API key needed

The MCP server does not call an LLM — your IDE does. It only reads your model and answers the IDE's questions, so none of the provider configuration on the Intelligence & AI page applies here.

Tools the server exposes¶

Read-only by default:

Tool	What it does
`memo_project_summary`	Element and relationship counts, per-layer completeness, error and warning totals
`memo_ontology`	Every legal element kind with its SysML construct and layer, and every legal relationship type
`memo_search_elements`	Find elements by name/id substring, kind, or layer
`memo_get_element`	One element in full: attributes, docs, source file, every relationship, its violations
`memo_trace`	Follow the relationship chain out of an element, to a given depth
`memo_validate`	Run the closure rules and return every violation

The two that change outcomes most are memo_ontology — which stops the agent inventing kinds your project does not define — and memo_validate, which lets it check its own edit rather than declaring success.

Read-only vs write¶

By default the server cannot modify your model. The agent reads, and writes SysML through your editor, where you review the diff as you would any other code change.

To let the agent create elements directly:

memo mcp init --write

That adds one more tool:

Tool	What it does
`memo_create_element`	Create an element and write it to the project SysML source

It still validates the kind against your ontology and refuses an id that already exists.

Consider whether you want this

Read-only is the default deliberately. In read-only mode every change to a regulated file arrives as an editor diff you approve. With --write, the agent writes directly, and your review moves to git diff. For a model under design control, read-only plus your normal review process is usually the right answer.

Calling a write tool on a read-only server returns a clear error telling the agent to ask you to restart with --write, rather than failing silently.

Running the server by hand¶

Useful when debugging a client that will not connect:

memo mcp --project .

The server speaks JSON-RPC on stdout and logs to stderr, so a bare run looks like it has hung — it is waiting for a client. Press Ctrl+D to end it.

Option	Default	Description
`--project <dir>`	current directory	Project to serve
`--write`	off	Enable the model-editing tool

The Cursor rules file¶

.cursor/rules/memo.mdc applies whenever Cursor works on a .sysml file. It tells the agent to check the ontology before editing, to use exact element ids, to treat terms drawn from a standard as regulated vocabulary rather than paraphrasing them, and to run memo_validate afterwards.

It is a normal file — edit it to suit your team's conventions. memo mcp init overwrites it, so keep local changes elsewhere or re-apply them after re-running init.

Troubleshooting¶

The server does not appear in Cursor. Restart Cursor after memo mcp init — it reads .cursor/mcp.json at startup.

"No MEMO project found." The --project path must contain the native entrypoint model/catalog/project.sysml. Run memo mcp init from the project root so the generated path is correct.

Answers look stale. The server reloads when your .sysml files change, checked per request. If an answer looks wrong, confirm the agent is not quoting an earlier reply from its own context rather than calling the tool again.