MCP Overview

The Model Context Protocol is a JSON-RPC standard for exposing tools to AI agents. PromptGate plugs into MCP three different ways depending on what you’re trying to do:

Surface	Role	URL	Project type
MCP Bridge	Expose your AI Gateway endpoints AS MCP tools to agents	POST /api/{uuid}/mcp	ai_gateway
MCP Gateway	Aggregate multiple upstream MCP servers under one URL	POST /api/{uuid}/mcp	mcp_gateway
MCP Control Plane	Manage PromptGate itself via MCP	POST /api/control/mcp	global

All three speak JSON-RPC 2.0. All three implement the standard initialize, tools/list, tools/call, ping, notifications/initialized methods. The endpoint URL determines which one you’re talking to.

Pick the right one

Bridge

You have an ai_gateway project with endpoints like summarize-ticket, classify-email, extract-fields. You want an agent (Claude Desktop, Cursor, your own) to discover and call these as tools.

→ Enable expose_as_mcp_tool on each endpoint, point the agent at /api/{uuid}/mcp, done.

Gateway

You’re using multiple upstream MCP servers — a filesystem server, a GitHub server, a Slack server. Instead of configuring each agent with N URLs, you want one URL that aggregates all of them.

→ Create an mcp_gateway project, register each upstream as an MCP Server, point the agent at /api/{uuid}/mcp. Tool names are namespaced (fs__read_file, gh__create_issue) so calls route deterministically.

Control Plane

You want an agent to manage PromptGate: list projects, rotate tokens, toggle endpoints, pull metrics.

→ Use the global /api/control/mcp with an admin-scoped token. 12 tools cover the read + write surface.

All three are bearer-authenticated

Every MCP surface requires a Bearer token:

Surface	Required scope
Bridge	mcp
Gateway	mcp
Control Plane	admin

See Client Tokens for token issuing.

All three speak the same protocol

The JSON-RPC envelope shape is uniform:

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/list"
}

{
  "jsonrpc": "2.0",
  "id": 1,
  "result": {
    "tools": [
      { "name": "summarize-ticket", "description": "...", "inputSchema": {...} }
    ]
  }
}

What changes between surfaces is which tools tools/list returns and what tools/call does:

• Bridge: tools = your AI endpoints; call = run the endpoint pipeline.
• Gateway: tools = union of upstream tools (with prefix); call = forward to the right upstream.
• Control Plane: tools = PromptGate management ops; call = read DB or mutate state.

Notifications + batch

All three surfaces support:

• Notifications (envelope without id) — return 204 No Content. Used by notifications/initialized.
• Batch requests (top-level array of envelopes) — return an array of responses (notifications omitted).

Compatibility

The PROTOCOL_VERSION constant is 2024-11-05 across all three surfaces — that’s the MCP spec version PromptGate implements.

Tested against: Claude Desktop, Cursor, custom JSON-RPC clients.

Where to go next

• MCP Bridge — exposing AI endpoints to agents.
• MCP Gateway — aggregating upstream MCP servers.
• MCP Control Plane — managing PromptGate via MCP.
• MCP API — the wire-format reference.

---

© Akyros Labs LLC. All rights reserved.