Sdcmodeler MCP server

sdcmodeler

<!-- mcp-name: io.github.Axius-SDC/sdcmodeler -->

The modeler's client for SDCStudio. Search the public catalog, estimate, assemble and publish an SDC4 data model, generate its package and an application. A Python client with httpx as its only dependency, a CLI, and sdcmodeler mcp, an MCP server for any agent.

Why

An agent that builds a feature decides a data model at that moment. With sdcmodeler the data is modeled in SDC by default: the agent searches the catalog first, reuses published components wherever one fits, mints only what is new, publishes the model, and takes the package. The result is interoperable because it is composed from shared, published components, and verifiable because every record validates against a published schema.

The name follows the role pattern of the other Semantic Data Charter servers: sdcvalidator validates, sdcgovernance governs, sdcreader reads, sdcmodeler models. It also names what SDCStudio calls every user: a Modeler.

Install

pip install sdcmodeler

The catalog and estimates need no account. Building needs an SDCStudio API key, created in Settings, API Keys, with the scopes and spending limit the holder should have:

export SDCSTUDIO_API_KEY=sdcs_...

The three layers

The client

from sdcmodeler import SDCStudioClient

c = SDCStudioClient()                       # key from $SDCSTUDIO_API_KEY, URL from $SDCSTUDIO_URL
hits = c.search_components("blood pressure", type="xdquantity")
tree = {
    "label": "Vitals",
    "components": [
        {"ct_id": hits["results"][0]["ct_id"]},            # reuse: free
        {"label": "Pulse", "data_type": "XdCount"},        # mint: 100 credits
    ],
    "clusters": [],
}
estimate = c.estimate(tree, "Vitals")       # the server's estimate: reuse, mint, problems, total_credits, wallet, key, estimate_id
result = c.assemble("Vitals", tree, estimate_id=estimate["estimate_id"])   # 500 credits plus the mints

Every leaf existing: the server assembles, publishes and generates the package in one call (result["mode"] == "reuse"; published and generated say so). Some leaves new: the pipeline runs and result["jobctid"] is polled with c.jobstatus(...), or c.waitfor_job(...), until completed, when the job has published and generated as asked and its result carries the model's links. publish=False leaves a draft for a person to review. The server refuses a build whose tree changed since the estimate.

A 402 is InsufficientCredits with the numbers in credits. A 403 from the key's policy (no spend scope, or past its limit) is KeyPolicyRefused. The client never funds a wallet; a person does that at the wallet page.

The CLI

sdcmodeler search "blood pressure" --type xdquantity
sdcmodeler models "patient"
sdcmodeler model <ct_id>
sdcmodeler artifact <ct_id> xsd --dir ./schemas
sdcmodeler estimate tree.json --title "Vitals"
sdcmodeler build tree.json --title "Vitals" --wait
sdcmodeler job <job_ct_id>
sdcmodeler package <ct_id> --dir ./packages
sdcmodeler app <ct_id> --profile lightweight --dir ./apps
sdcmodeler wallet

The MCP server

sdcmodeler mcp

A JSON-RPC 2.0 stdio server, no SDK, the way sdcvalidator and sdcgovernance are built. A client configuration needs only:

{"command": "uvx", "args": ["sdcmodeler", "mcp"], "env": {"SDCSTUDIO_API_KEY": "sdcs_..."}}

Shortened. The full README is on GitHub.