ModelBrain
Exhibit F · Obligations of the agentMCP server · HTTP transport

The agent shall call can_i_use before from_pretrained.

ModelBrain is an MCP server. Your agent calls it before it writes from_pretrained, and gets the verdict, the lineage path and the clause straight from FalkorDB.

§1

Claude Code

claude mcp add · one command
claude mcp add --transport http modelbrain https://api-production-fc09a.up.railway.app/mcp

Then ask: "add OPEA/Llama-3.2V-11B-cot-int4-sym-inc to our pipeline for our German product". The agent calls can_i_use and reports BLOCKED with the path.

§2

Cursor and other MCP clients

mcp.json · mcpServers
{
  "mcpServers": {
    "modelbrain": {
      "url": "https://api-production-fc09a.up.railway.app/mcp"
    }
  }
}

§3

Tools

10 tools, each answered by one graph query.

ToolWhat it answers
can_i_useVerdict for a model given HQ country, markets and commercial use, with path and clause.
trace_weightsThe true parent from ~2 KB of sampled weights, merges, and the obligations the weights carry.
explainEvery ancestry path, the clauses with quotes and sources, and trust tiers.
lineageAncestors and descendants with declared licences and trust tiers.
proof_pathShortest derivation path to a given base model, or proof that none exists.
relabel_hubsPageRank over lineage: models whose card is permissive but whose ancestors are not.
chokepointsBetweenness: models where a licence fix reaches the most descendants.
familiesLabel propagation: model families in the lineage graph.
confirm_edgeA company confirms a lineage edge; two independent confirmations raise its trust tier.
watchSubscribe a company to upstream changes; a change reopens its rulings.

§4

In CI

(b) The actions/modelbrain-check GitHub Action scans a pull request for Hugging Face model ids, asks the same graph, writes a job summary with path and blocking clauses, and fails the check when a model is BLOCKED for your headquarters.

Automated screening, not legal advice.