@trstlyr/mcp

MCP (Model Context Protocol) server for TrstLyr Protocol — gives AI agents the ability to check trust before acting.

Install it in Claude Desktop, Cursor, or any MCP-compatible host in minutes.

Tools

| Tool | Description | |------|-------------| | trust_query | Full trust report: score, risk level, signals, and evidence | | should_proceed | Binary go/no-go check with reasoning | | trust_explain | Narrative explanation of why a subject has its trust rating |

Quick Install (Claude Desktop)

1. Clone and build:

git clone https://github.com/tankcdr/aegis.git
cd aegis
pnpm install
pnpm -r build

2. Add to Claude Desktop config (~/Library/Application Support/Claude/claude_desktop_config.json):

{
  "mcpServers": {
    "aegis": {
      "command": "node",
      "args": ["/path/to/aegis/packages/mcp/dist/index.js"],
      "env": {
        "GITHUB_TOKEN": "ghp_your_token_here"
      }
    }
  }
}

3. Restart Claude Desktop. Three new tools appear automatically.

Subject Format

Use namespace:id format:

| Subject | Meaning | |---------|---------| | github:tankcdr | GitHub user | | github:tankcdr/aegis | GitHub repository | | tankcdr/aegis | Shorthand — defaults to github namespace |

Example Prompts

"Before I run this, check if github:some-user/some-script is trustworthy."

"Should I install github:modelcontextprotocol/servers? Check trust first with action install."

"Explain why github:tankcdr/aegis has its trust rating."

"Check trust for openai/openai-node — I'm about to run it."

Environment Variables

| Variable | Description | |----------|-------------| | GITHUB_TOKEN | GitHub personal access token. Increases rate limit from 60 to 5,000 req/hr. Optional but recommended. |

How It Works

Each trust query runs through the Aegis 7-step pipeline:

1. Identity resolution — parse subject into canonical namespace:id
2. Signal dispatch — fan out to all eligible providers in parallel
3. Fraud detection — lightweight anomaly detection
4. Subjective Logic fusion — cumulative belief fusion (Jøsang, 2001)
5. Ev-Trust adjustment — λ=0.15 evolutionary stability penalty on conflicting signals
6. Risk mapping — score → risk level → recommendation
7. Cache — results cached in-memory (TTL: 5 min default)

License

Apache 2.0