Agntor MCP

What agntor/mcp MCP server does

The agntor/mcp MCP server exposes Agntor’s agent trust and administration functions to MCP-compatible clients. Its tools cover four main areas: discovering and evaluating agents, guarding agent interactions, supporting agent-to-agent payments, and managing registered agents.

For discovery and certification, clients can retrieve an AgentCard, fetch an EIP-8004-compatible registration file, inspect health and behavioral metrics, check certification status, calculate a trust score, register an agent, and trigger identity verification through the SDK. The security tools scan incoming prompts for prompt injection and unsafe instructions, redact PII or secrets from outputs, and apply allow or deny policies to tool execution.

Payment-related operations include creating escrow tasks and issuing signed JWT audit tickets for x402 transactions. Administrative tools can search agents by trust score, tier, or capabilities and activate an emergency kill switch for an agent.

How it works

The agntor/mcp MCP server communicates with clients through the MCP protocol using either stdio or HTTP. The README describes support for clients including Claude Desktop, Cursor, and Cline, and also provides a hosted MCP endpoint.

The server uses local SDK utilities for guarding input, redacting output, and enforcing tool policies. Trust, identity, discovery, and related backend operations call the Agntor REST API through the SDK. Applications can also import createAgntorMcpServer and connect the resulting server to an HTTP, stdio, WebSocket, or another transport.

Setup and configuration

Install the package globally with npm or run it directly from an MCP client using npx -y @agntor/mcp. The client configuration uses npx with the package name and the -y flag. Standalone stdio mode is available with the package’s --stdio option.

HTTP operation uses AGNTOR_API_KEY and AGNTOR_SECRET_KEY in the documented startup example. AGNTOR_API_KEY is required for API tools. AGNTOR_MCP_AUTH_KEY protects the HTTP MCP endpoint and falls back to the API key when it is not set. The backend URL defaults to https://app.agntor.com, and the HTTP server defaults to port 3100. MCP_TRANSPORT can force the transport mode, including stdio.

The hosted endpoint is https://mcp.agntor.com/mcp. When authentication is enabled there, requests include the X-AGNTOR-API-KEY header.

Tools and capabilities

The 14 exposed tools include:

• Agent identity and trust: get_agent_card, get_agent_registration, check_agent_pulse, is_agent_certified, get_trust_score, register_agent, and verify_agent_identity.
• Safety controls: guard_input, redact_output, and guard_tool.
• Commerce and audit: create_escrow and issue_audit_ticket.
• Administration: query_agents and activate_kill_switch.

The agntor/mcp MCP server therefore fits workflows that need to inspect an agent before delegation, screen model inputs and outputs, authorize tools, create payment arrangements, or disable an agent during an incident.

Limitations and notes

API-backed operations depend on Agntor credentials and its backend service. The README does not specify pricing, service quotas, certification criteria, or the detailed behavior of the trust-score calculation. HTTP authentication is described as conditional, while the API key is explicitly required for API tools. The repository is licensed under MIT.