AI Security Gateway with A2A (Agent-to-Agent) protocol support, OAuth authentication, MCP proxy integration, and agent registration/invocation capabilities
This MCP server exhibits critical definition quality failures across nearly all dimensions. While 8 tools are declared with some schema structure, the schemas are severely incomplete (parameter types often missing or inferred), descriptions range from superficial to actively misleading, and error handling is entirely absent. The server appears to expose a custom JSON-RPC gateway API (agents/invoke, oauth-proxy/token, etc.) rather than implementing proper MCP tool patterns. Parameter descriptions lack actionable guidance; schemas mix JSON-RPC protocol boilerplate ('jsonrpc', 'method', 'id') with actual tool logic, violating separation of concerns. No tool declares output schemas or error recovery guidance. Risk annotations exist (READ_ONLY, WRITE) but are not tied to any error-handling or permission-gating patterns. The examples show curl/bash/Go client code rather than tool definitions, suggesting the MCP server itself is not properly instrumented. Overall, this reads more like an API documentation dump than a production MCP tool registry.
Invoke an A2A agent via Gateway JSON-RPC with optional streaming responses
List all accessible A2A agents
Register a new A2A agent in the AI Security Gateway via URL or AgentCard JSON
Call a specific MCP tool with given arguments via OAuth-authenticated proxy
List available MCP tools via the OAuth-authenticated MCP proxy
Dynamic Client Registration (DCR) for OAuth Proxy - register a new OAuth client
Exchange authorization code for access token using PKCE
Output schemas entirely missing. No tool declares what it returns to the agent, only examples in bash/Go client code. LLMs cannot plan downstream calls or extract required fields without documented return types.
Tool descriptions are superficial and lack context for LLM selection. Examples: 'Invoke an A2A agent via Gateway JSON-RPC with optional streaming responses' (59 chars) and 'List all accessible A2A agents' (30 chars). None explain WHEN to call, WHAT the result contains, or what happens on error.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 37 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 23 | - | v1 |
Start OAuth authorization flow for a given provider
Parameter types missing or incomplete. 'agents/invoke' message parameter is typed as 'object' with nested properties, but no 'required' array, minProperties/maxProperties, or validation rules. 'oauth-proxy/register' grant_types and response_types are arrays but lack itemType specifications.
Parameters mix protocol boilerplate with tool logic. 'mcp/tools/list' and 'mcp/tools/call' expose 'jsonrpc', 'method', and 'id' as input parameters, these are JSON-RPC envelope details, not tool-level concerns. This violates separation of concerns and forces LLMs to reason about protocol internals.
No error handling or recovery guidance. None of the 8 tools document what errors they can return, which are retryable, or how to fix them. No examples of error responses provided. Agents cannot self-correct on failure.
'oauth-proxy/register' accepts 'client_secret' as a tool input parameter. Secrets must NEVER be exposed as tool parameters, they get logged in agent traces and could leak into prompt history. Use server-side secret injection instead.
Parameter descriptions lack specificity and actionable guidance. Example: 'agents/register' -> 'agent_card': 'AgentCard JSON definition (alternative to URL)' does not explain what fields are required in the JSON, what format, or constraints. Similarly, 'oauth-proxy/register' -> 'scope': just 'OAuth scopes to request (optional)' with no examples of valid scopes or format.
Tool naming uses slash separators ('agents/invoke', 'oauth-proxy/register') which are uncommon in production MCP tools and may cause parsing issues in some LLM contexts. Standard naming is verb_noun (e.g., 'invoke_agent', 'register_oauth_client') or namespace_verb_noun (e.g., 'agents_invoke_agent'). Slash separators conflate HTTP path structure with tool identity.
Risk annotations ('READ_ONLY', 'WRITE') are declared but not tied to tool annotations in the MCP schema (readOnlyHint, destructiveHint, idempotentHint). Modern MCP expects structured hints in the tool definition, not external risk labels.
'oauth-proxy/register' has mutually exclusive parameters ('url' vs 'agent_card' in agents/register, 'client_secret' context in oauth-proxy/token) but does not document the mutual exclusivity. LLMs will pass both, causing ambiguous or failing calls.