Amazon Ads MCP Token Cost: Compare Tool Catalogs

Amazon Ads MCP token cost starts with the tool definitions loaded at connection. Amazon's official Ads MCP Server hands your agent 92,678 tokens of tool definitions the moment it connects, before it reads a single campaign. Kuudo's Amazon Ads MCP hands it 388, with 756 Amazon Ads operations behind them. This guide covers what that gap is, why it repeats, and why it widens as a catalog grows.

Amazon Ads MCP is a Model Context Protocol server that exposes Amazon Ads API operations as tools an AI client such as Claude, ChatGPT or Cursor can call, so the assistant reads and changes campaigns through Amazon's API instead of a pasted report. Amazon's official server has been in open beta since February 2026; Kuudo's is open source and built directly on the same API. Other vendors' servers are in the comparison of Amazon Ads MCP servers.

Tokens are the meter on agent workloads: API models are priced per million of them, and every window has a ceiling. Conversation history, retrieval payloads and output length vary from run to run. A server's advertised tool catalog does not. It is a fixed cost that lands in the first request of a session and, on clients that inject the full tool list, is re-sent on every turn whether or not a tool is called.

The Amazon Ads MCP catalog is charged before your agent does any work

When an MCP client asks a server for its tools, the server answers with the definition of every tool it advertises: a name, a description, and a JSON schema for the inputs. On clients that load the full list, that answer sits in the context window from the first request on, and its size is a design decision the server made before you arrived.

Amazon's largest tool, the one that creates a target, is 12,051 tokens on its own: 31 times Kuudo's entire advertised surface. Schema makes up 94% of Amazon's catalog, ten tools carry more than half of it, and sixteen exceed 1,000 tokens each. Kuudo's Amazon Ads MCP advertises three small tools, so the MCP connection itself weighs almost nothing.

Everything useful arrives after the connection and competes for the same room: the campaign data a tool returns, what Amazon Agent Atlas retrieves when an Amazon rule applies, and the judgment a Skill carries into the run. Those are the costs worth paying. The catalog is not.

On clients that resend it, the MCP catalog repeats every turn and multiplies across sub-agents

Clients that inject the full tool list send the advertised definitions on every inference call, so the connect cost is not a one-time charge that frees up. It is in every request for the life of the session. Clients that load definitions on demand pay only for what they fetch. Where the catalog is resent, two consequences follow that a bigger window does not fix.

Sub-agents multiply it. Each sub-agent context carries the catalog on its own, so ten parallel sub-agents on Amazon's server carry 926,780 tokens of tool definitions between them. On Kuudo they carry 3,880.

The constraint changes with the deployment tier and does not go away. On 128k and 200k class models, including open-weight and self-hosted deployments, 92,678 tokens is 72% and 46% of the window before any work, and capacity is the limit. On Claude models with 1M windows the catalog fits at 9%, and price takes over. Carrying it uncached across a 20-turn session, estimated by pricing the cl100k_base counts at Anthropic's published base input rates:

Claude's own tokenizer counts text differently from the cl100k_base tokenizer used here, and a different count moves both columns. Prompt caching cuts both columns by the same multiplier, because caching lowers the price of re-sending tokens, not the number of tokens sitting in the window. Cheap tokens, no room. Plenty of room, expensive tokens. A fixed catalog loses on whichever axis your deployment cares about.

Research on long contexts adds a third cost. Chroma's context-rot study found model performance degrades as input length grows, with distractors making it worse, and the Lost in the Middle study found models miss information placed in the middle of a long input. Both are in the sources.

Kuudo's Amazon Ads MCP catalog grows without adding a token at connect

Under a static catalog, advertised breadth and connect cost are the same number. Every tool added is charged to every client in every session, called or not, at about 842 tokens on Amazon's current per-tool average. Kuudo separates the two. Its Amazon Ads MCP reaches 756 Amazon Ads API operations at a connect cost of 388 tokens, and the 757th adds nothing at connect.

What happens next: connect the lighter Amazon Ads MCP

Connect the Kuudo Amazon Ads MCP. It is the lighter server by every measure in this guide: 388 tokens at connect against 92,678, and a catalog that costs nothing at connect to grow. The window holds your campaigns instead of a catalog.

Its 756 operations reach Sponsored Products, Brands, Display and TV, Amazon DSP, Amazon Marketing Cloud (AMC) workflows and audiences, Attribution, Marketing Stream and reporting. Amazon's published tool groups do not list AMC audiences, Attribution or Marketing Stream. The comparison with Amazon's official server lays it out row by row, with every Amazon fact cited to Amazon's own pages.

The same client can hold the Selling Partner MCP and the Vendor Central MCP beside it. Inventory, orders, purchase orders and bids sit in one session, which is the whole Amazon job rather than the advertising slice of it.

• Your cloud, your Amazon app. The servers run in your own cloud on your own Amazon app, so calls go to Amazon and not through a vendor.
• Your policy on every write. Reads return data; writes wait on the policy you set, and every call is on the record.
• Your workflows as Skills. Skills turn a workflow into a playbook your team can read and run from Claude, ChatGPT, Cursor, Codex, n8n or any client that speaks MCP.
• Your results out of the window. The Amazon Agent Data Layer returns a bounded preview and a result handle for a large query, so the context holds the evidence and not the whole table.

The method and the open items are in the reference layer below.

Next: why the same AI in every competitor's hands makes your account context the edge.