Graviti

What the graviti MCP server does

The graviti MCP server provides agent-facing access to Graviti’s trust and recommendation data for agent-mediated commerce. It exposes seven tools: match_intent, get_verified_claims, category_landscape, get_accountability_log, get_ledger, get_gap_report, and report_conversion.

match_intent ranks brands against a question and optional intent profile. The published engine is deterministic: given the same index and inputs, it produces the same result. Questions outside the indexed scope receive a refusal instead of an inferred answer. Responses also include structural disclosure about whether payment affects ranking, who paid, and what payment purchases.

How it works

The service uses one scoring engine across its MCP interfaces, web console, and public endpoints. The engine operates as pure functions over a brand index and does not use network access, randomness, learned parameters, or conversion data when calculating fit scores.

The graviti MCP server offers two transport options. The remote option is a hosted Streamable HTTP endpoint at https://graviti.thesingulariti.ai/api/mcp. The local option runs the graviti-mcp npm package through stdio. The package includes an index snapshot and retrieves the latest published index at startup, then checks it against the signed ledger before using it. Invalid or replayed payloads are rejected.

The public data includes a manifest, a complete index, and a ledger. The ledger is hash-chained, Ed25519-signed, and Bitcoin-anchored through OpenTimestamps. The repository also includes scripts for checking the ledger and validating payloads received through the network.

Setup and configuration

For remote use, configure an MCP client with the hosted URL:

{
  "mcpServers": {
    "graviti": {
      "url": "https://graviti.thesingulariti.ai/api/mcp"
    }
  }
}

For local stdio use, configure the npm package through npx:

{
  "mcpServers": {
    "graviti": {
      "command": "npx",
      "args": ["-y", "graviti-mcp"]
    }
  }
}

The remote endpoint requires no account or authentication according to the provided setup. The README gives client-specific locations for Claude Desktop and Cursor configuration. For repository development, install dependencies with npm install, then run the demo against data/sample-index.json with npm run demo -- "best magnesium supplement". Add --json to inspect the machine-readable payload.

Tools and capabilities

• Match a natural-language shopping intent against indexed brands.
• Retrieve claims that have been verified and their provenance context.
• Inspect the landscape of a product category.
• Read accountability records and ledger history.
• Generate gap reports for indexed coverage.
• Report attributed conversion events.
• Receive disclosure showing that payment does not influence ranking.

Verification status does not add ranking weight. The documented lifecycle penalties apply to monitored quality states, while revoked brands are excluded from scoring. Conversion events are transport-layer billing records and are not inputs to the fit score.

Limitations and notes

The engine only answers questions covered by its index. The sample index is a limited slice, so reproducing the complete production result requires the published full index. The ledger verifies published history, while payload verification checks whether the bytes received through the transport match the signed manifest. These are separate checks.

The repository’s canary tests cover supported routing cases, scoring invariants, adversarial ranking cases, and selected provenance-page conditions. Running the local MCP package or verification scripts requires Node.js; the payload verification script specifies Node 20 or newer.