MCP (Model Context Protocol)¶

MCP is the primary agentic interface for the seller agent. Buyer agents call seller tools directly through structured, typed tool calls over MCP. This is the preferred protocol for automated workflows where the buyer agent knows exactly which operation it needs.

Endpoints¶

Endpoint	Method	Description
`/mcp/sse`	GET	Streamable HTTP transport (SSE) for persistent tool sessions
`/mcp/tools`	GET	Tool discovery --- lists all available tools and their schemas
`/mcp/call`	POST	Execute a tool call with typed arguments

Available Tools¶

The seller agent exposes the following tools via MCP. Each maps to a core seller operation:

Tool	Description
`list_products`	Browse the product catalog with optional filters
`get_product`	Get full details for a specific product
`get_pricing`	Tiered pricing calculation with buyer context (tier, volume, deal type)
`check_availability`	Check inventory avails for a product, date range, and impression volume
`create_proposal`	Submit a proposal for seller review
`create_order`	Create an order from an accepted proposal
`create_line_item`	Add line items to an existing order
`book_programmatic_guaranteed`	Book a Programmatic Guaranteed (PG) deal directly
`create_pmp_deal`	Create a Private Marketplace (PMP) deal and return a Deal ID
`get_deal_status`	Check the status of an existing deal
`request_quote`	Request a non-binding price quote (IAB Deals API)
`book_deal`	Book a deal from an accepted quote (IAB Deals API)

Example Tool Call¶

{
  "name": "get_pricing",
  "arguments": {
    "product_id": "premium-video",
    "buyer_tier": "agency",
    "volume": 5000000
  }
}

The response is a typed JSON object with the pricing breakdown, including base CPM, tier discount, volume discount, and effective CPM.

When to Use MCP¶

MCP is the right choice when the buyer agent:

Knows exactly which operation to perform (e.g., check availability, book a deal)
Needs deterministic, structured responses
Is running an automated workflow without human intervention
Wants the fastest possible execution path

Protocol Comparison¶

	MCP	A2A	REST API
Interface style	Structured tool calls	Natural language (JSON-RPC)	HTTP request/response
Best for	Automated agent workflows	Discovery, negotiation, complex queries	Human operators, dashboards
Response format	Typed tool results	Mixed text + structured data	JSON
Multi-turn	Stateless per call	Yes, via `contextId`	Stateless
Speed	Fastest	Moderate (LLM processing)	Fast
Determinism	Fully deterministic	Non-deterministic (LLM)	Fully deterministic
Transport	SSE (Streamable HTTP)	HTTP POST (JSON-RPC 2.0)	HTTP verbs

MCP Server Examples¶

The repository includes two MCP server examples:

examples/publisher_gam_server.py --- Publisher seller agent with Google Ad Manager integration. Exposes CTV inventory via MCP, books PG lines in GAM, and creates PMP deals.
examples/dsp_server.py --- DSP seller agent. Receives Deal IDs from buyers, activates deals on campaigns, and books performance and mobile app campaigns.