MCP serverLive · 10 tools

Evidence layer for AI agents, native to Claude and Cursor

Connect any MCP-compatible runtime — Claude Desktop, Claude Code, Cursor, VS Code — directly to CertifiedData. Every tool is typed, every mutating call returns a machine-verifiable proof artifact (signed receipt, rotated token, or attestation challenge) that third parties can verify without trusting our servers.

Current MCP scope

This MCP server currently focuses on commerce, agent identity, dataset licensing, and receipt verification.

Dataset generation and decision logging are first-class capabilities of CertifiedData, but are exposed today via the HTTP API and the TypeScript / Python SDKs rather than through MCP. This keeps the MCP surface narrowly scoped to what agents typically do in-loop: discover products, check policy, transact, verify, and manage access. Tool coverage will expand over time; track changes in the changelog.

Install — sandbox

Add this to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%/Claude/claude_desktop_config.json (Windows).

Claude Desktop — sandbox

{
  "mcpServers": {
    "certifieddata-sandbox": {
      "command": "npx",
      "args": ["@certifieddata/mcp-server"],
      "env": {
        "CDAC_API_KEY":  "cdp_test_...",
        "CDAC_BASE_URL": "https://certifieddata.io",
        "CDAC_ENV":      "sandbox"
      }
    }
  }
}

Claude Desktop — live

{
  "mcpServers": {
    "certifieddata-live": {
      "command": "npx",
      "args": ["@certifieddata/mcp-server"],
      "env": {
        "CDAC_API_KEY":  "cdp_live_YOUR_KEY_HERE",
        "CDAC_BASE_URL": "https://certifieddata.io",
        "CDAC_ENV":      "live"
      }
    }
  }
}

Exposed tools — 10 live

Group	Tool	Arguments	Returns	Trust
Commerce	browse_marketplace	{}	List of certified datasets with slug, price, rows × columns, algorithm, cert signer	Read-only. No side effects.
Commerce	check_spend_policy	{}	Active policy: rails, max_per_tx_cents, daily_limit_cents, merchant allow/block lists	Read-only.
Commerce	check_remaining_budget	{}	Today's spend, daily limit, remaining budget	Read-only.
Commerce	authorize_purchase	{ dataset_slug, amount_cents, buyer_email, note? }	authorization_id, status, policy_decision	Writes an authorization record. No money moves.
Commerce	execute_purchase	{ authorization_id, buyer_email?, dataset_slug? }	receipt_id, download_url, license_key, dataset_hash, certificate_url, verify_url	Charges the configured rail. Ed25519-signed receipt issued.
Receipts	verify_receipt	{ receipt_id }	valid, signatureValid, hashValid, signingKeyId, receipt_payload	Read-only. Third party can verify via /api/payments/verify/{receipt_id}.
Identity	get_agent_identity	{ agent_id }	DID, display_name, status, key set, verified attestations	Read-only.
Identity	verify_agent_attestation	{ agent_id, purpose?, key_id? }	challenge nonce, attestation_id, expires_at	Issues a one-time challenge. Client signs offline and posts to /attest.
Licensing	get_license	{ license_id }	dataset_slug, SPDX, usage rights, commercial/redistribution flags, status, verify_url	Read-only.
Licensing	rotate_license_token	{ license_id, reason? }	Fresh download_url valid for 48 hours, expires_at	Mints a new time-limited download token. Existing token is superseded.

Demo prompt

Paste this into Claude Desktop with the MCP server connected. Expected runtime: 15–30 seconds.

Full browse → policy → authorize → execute → verify flow

You are a data procurement agent for an AI research lab.

Your task:
1. Browse the CertifiedData marketplace and find a synthetic dataset suitable
   for training a customer support classification model.
2. Check your payment policy to confirm the purchase is within your authorized
   spend limits.
3. Check today's remaining budget.
4. Authorize the purchase for the dataset you selected.
5. Execute the purchase. Report the full delivery bundle: download URL,
   license key, SHA-256 hash, certificate URL, and verify URL.
6. Verify the receipt cryptographically and confirm: valid=true,
   signatureValid=true, hashValid=true.

Safety model

• The MCP server is a thin RPC layer — all policy evaluation, signing, and rail execution happens server-side at certifieddata.io.

• Prompt-injected tool arguments cannot bypass spend limits, policy rules, or signature validation — these are enforced server-side against the authenticated API key.

• A sandbox key can be rotated or revoked at any time via /dashboard/api-keys.

• Full trust boundary model →

SDK Agent Commerce Error taxonomy Rate limits agent.json