Agent Mandate MCP

Verification-only MCP server exposing one tool: verify_action.

The tool sends caller-supplied mandate and action facts to Agent Mandate's POST /v1/verify endpoint and returns its allow, deny, or requires_approval receipts. It does not issue or revoke mandates, execute an action, change an account, or call billing. It makes at most one API request per tool call and never retries automatically.

Status: public source; npm package and Official MCP Registry release pending. The repository link below is live. The npm and Registry identities become available only after their separate public releases:

• npm: @api-disk-integrations/agent-mandate-mcp@0.1.0

• Official MCP Registry: io.github.API-Disk-Integrations/agent-mandate@0.1.0

• source: API-Disk-Integrations/agent-mandate-mcp

Install and run

After the exact package version is publicly published and independently read back, run it with the version pinned:

AGENT_MANDATE_API_KEY="$AGENT_MANDATE_API_KEY" \
  npx --yes @api-disk-integrations/agent-mandate-mcp@0.1.0

Do not commit a literal credential. A generic stdio client configuration is:

{
  "mcpServers": {
    "agent-mandate": {
      "command": "npx",
      "args": ["--yes", "@api-disk-integrations/agent-mandate-mcp@0.1.0"],
      "env": {
        "AGENT_MANDATE_API_KEY": "${AGENT_MANDATE_API_KEY}"
      }
    }
  }
}

${AGENT_MANDATE_API_KEY} denotes the host's secret/environment reference. Use the client host's documented secret facility if its interpolation syntax differs. The package uses stdio and reads exactly that environment variable; it has no remote /mcp endpoint.

Related MCP server: @palveron/mcp-server

Tool contract

There is exactly one tool, verify_action. Supply either one action or an actions array, never both. A batch contains 1–500 actions. Monetary amounts are non-negative integer minor units and require currency.

Example input:

{
  "mandate": {
    "mandate": {"id": "mnd_example"},
    "signature": "caller-supplied-signature"
  },
  "action": {
    "agent": "procurement-agent",
    "action": "payments.transfer",
    "resource": "vendor.acme",
    "amountMinor": 2500,
    "currency": "USD",
    "at": "2026-09-05T20:00:00Z"
  }
}

Representative successful structured output:

{
  "count": 1,
  "receipts": [
    {
      "decision": "deny",
      "mandateId": "mnd_example",
      "violations": [{"code": "action_not_granted", "detail": "No matching grant"}]
    }
  ]
}

Treat all three decisions literally. In particular, neither allow nor requires_approval executes the proposed action.

Errors, limits, and safety

• If AGENT_MANDATE_API_KEY is absent, the tool returns an MCP error before making a request.

• A provider error is reduced to its numeric HTTP status and an allowlisted code. Provider-controlled message and request-ID text is never returned to the model.

• One tool call produces at most one HTTPS request to the fixed https://agentmandate-api.com/v1/verify endpoint. There is no automatic retry; Fetch uses explicit redirect: error, and redirect responses are rejected without a follow-up request.

• The serialized UTF-8 request and decoded response body each have a hard 1 MiB (1,048,576-byte) limit. The request is measured before Fetch. The response is counted while streaming before JSON parsing, regardless of a missing, misleading, or chunked Content-Length representation.

• The request timeout defaults to 30 seconds. A batch is also limited to 500 actions. Provider account rate and monthly usage limits still apply; consult the current product pricing and documentation before production use.

• The server is read-only with respect to mandate, account, and billing state, but an API verification consumes the caller's metered allowance.

• Node.js 20 or newer is required. This release is tested for MCP protocol revision 2026-07-28 and retains the SDK's listed 2025 compatibility revisions.

Links

• Product, pricing, privacy, and terms

• Developer documentation

• OpenAPI reference

• Service status

• Source repository

• Issue support

• Security reporting

• MIT license

• Threat model

The source link is not evidence that the npm package or Registry listing is available. Those two releases require their own public readback. A clone, install, download, tool call, or listing is not evidence of customer activation or revenue.