Table of contents

• Where this fits in Kuudo
• What are MCP tools?
• What is Amazon Ads API MCP SDK?
• Quick start
• Installation
• Configuration (auth, packages, profiles, regions)
• Downloading reports and exports
• Example MCP client (Claude Desktop)
• Context limits
• Code mode
• Tool audit
• Background tasks
• Frequently asked
• Troubleshooting
• Documentation map

Where this fits in Kuudo

This repository is the tool layer of a broader Amazon agent platform. Kuudo gives Amazon agents tools, knowledge, workflows, and a place to work. This repository is the tools.

Read this repository as one layer of that platform rather than as the whole of it. The operations below are what an agent can call; Skills are what it runs.

What Are MCP Tools?

Think of MCP (Model Context Protocol) as a translator between an AI model and outside systems (like Amazon Ads). Each MCP tool is like a remote control button that tells the AI how to interact with Amazon Ads. Without MCP tools, the AI would have no idea how to “talk” to Amazon Ads.

With MCP tools:

• The AI knows the exact endpoints to call.
• The AI can request campaign reports, budgets, or targeting data safely.
• Everything is structured, so the AI doesn’t break things by making random guesses.

👉 In short: MCP tools = a safe, well-labeled toolkit that lets AI work with the Amazon Ads API.

🚀 What is Amazon Ads API MCP SDK?

The Amazon Ads API MCP SDK is an open-source implementation that provides a robust foundation for creating AI-powered advertising tools, chatbots, and automated services.

✨ Key Features

• 🔌 MCP Integration: Full Model Context Protocol compliance for AI application integration
• 🌍 Multi-Region Support: NA, EU, and FE region endpoints with automatic routing
• 📊 Comprehensive API Coverage: Campaigns, profiles, reporting, DSP, AMC workflows, and more
• 📝 Type Safety: Full Pydantic model support with comprehensive type hints
• 🧪 Production Ready: Includes testing, validation, and error handling

🎯 Use Cases

Claude Desktop Integration

• Campaign Management: Ask Claude to create, update, or analyze campaigns
• Performance Insights: Get AI-powered analysis of your advertising performance
• Budget Optimization: Let Claude suggest budget adjustments based on performance
• Creative Testing: Get recommendations for ad creative improvements
• Reporting: Generate custom reports and insights on demand

AI Applications

• Marketing Chatbots: Build conversational AI that can manage Amazon Ads campaigns
• Automated Reporting: AI-powered insights and performance analysis
• Smart Budget Management: Intelligent budget optimization using AI
• Creative Optimization: AI-driven ad creative testing and optimization

Enterprise Services

• Marketing Automation Platforms: Integrate Amazon Ads into existing marketing tools
• Agency Management Systems: Multi-client, multi-account advertising management
• E-commerce Integrations: Connect Amazon Ads with e-commerce platforms
• Analytics Dashboards: Real-time advertising performance monitoring

Developer Tools

• API Wrappers: Create custom SDKs for specific use cases
• Testing Frameworks: Automated testing for Amazon Ads integrations
• Development Tools: Local development and debugging utilities

Why this Ads MCP is different

Other Amazon Ads MCPs do happy path. They fall apart the moment an agent makes a reasonable mistake and Amazon's Ads API has more reasonable mistakes than almost any other large API surface. The v1 reporting catalog uses a different vocabulary than v3. The docs are scattered across migration generations. The same field can have three plausible names depending on which tutorial an agent or human was trained on.

You've felt this before

You know the symptoms even if you've never thought of them as a single problem:

• "Compacting conversation…" right when you were getting somewhere
• "Usage limit reached" mid-task, with nothing to show for the tokens
• "Context window exceeded" start over, re-explain everything
• The agent quietly forgetting the fix it figured out three turns ago
• Five tool calls, four of them retries, none of them progress
• Re-pasting the same error into the chat for the third time

That's not the model being dumb. That's the API surface burning context vague errors trigger thrashing, thrashing fills the window, the window fills and everything useful gets evicted. Anywhere from 4 to 12 round-trips for an agent: wrong field names, wrong body shape, wrong filter operator, wrong date placement, wrong advertiser-account format.

Here, the wrong moves return the right move

This server is built around a different premise: the error surface is the documentation surface. Every failure is a teaching opportunity, and the server is engineered to make the next attempt smarter than the last one, for the same agent or the next one, without retraining, prompt updates, or memory tricks.

• The server tells you what went wrong, and why you got it wrong
• Every failure converts to a corrective action — the model just reads the hint and moves on
• A five-tool-call debugging arc collapses into a two-call corrected one
• The same mistake a week later doesn't re-burn context, because the validator, alias table, and deprecated-shape table are still right there
• No out-of-context tribal knowledge to forget

Every failure is cheaper than the last. Every agent learns from the same authoritative source. Every retry has direction. That compounding is what a smart agent solution looks like in practice.

Everything is an edge case, you can help

We won't catch every edge case. Amazon's API surface is enormous, the migration history is messy, and real-world failures get more creative than anything anyone can anticipate. What we're committing to is the strategy that errors should teach, that documentation should live in the surfaces agents actually touch, that every wrong move should make the next move easier.

If you hit a failure where the envelope didn't help; vague hint, wrong suggestion, no hint at all, that's exactly the feedback we want. Open an issue, paste the envelope, tell us what you expected. The strategy is only as good as the cases it covers.

Quick start

Prerequisites: Docker (recommended), Python 3.10+ if you run from source, and Amazon Ads API access (your own developer app or a partner such as Openbridge).

1. git clone https://github.com/KuudoAI/amazon-ads-mcp.git && cd amazon-ads-mcp
2. cp .env.example .env and add credentials (see Configuration).
3. docker compose up -d
4. Connect your MCP client to http://localhost:<PORT>/mcp/ — .env.example sets PORT=9080 (override in .env if needed).

OAuth steps, client JSON samples, and a full variable reference: Configuration below, INSTALL.md, and AGENTS.md.

📚 What Is Included In the Amazon Ads MCP?

The MCP server mirrors broad coverage of the Amazon Ads API surface. Each enabled package maps to a set of API operations. That includes Campaign Management (Amazon Ads API v1), Exports, Amazon Marketing Cloud, and many more.

Here is a representative list of the various Amazon API services in the MCP:

• Accounts
• Audiences
• Reporting
• Brand metrics
• Sponsored Products
• Sponsored Brands
• Sponsored Display
• Amazon DSP
• Amazon Attribution
• Recommendations & insights
• Creatives
• Change history
• Data provider
• Products
• Unified pre-moderation
• Moderation
• Amazon Marketing Stream
• Locations
• Exports
• Media Planning
• Amazon Ads API v1

🧪 Amazon Ads API v1

The Amazon Ads API v1 represents a reimagined approach to the Amazon Ads API, built from the ground up to provide a seamless experience across all Amazon advertising products through a common model. One major benefit of this common model is improved compatibility with code generation tools such as client library generators.

⚠️ Beta Notice: These APIs are currently in beta at Amazon. Features and endpoints may change. Use in production with caution.

To activate Ads API v1 packages, add them to your AMAZON_AD_API_PACKAGES environment variable:

# Example: Enable the merged Ads API v1 ALL surface
AMAZON_AD_API_PACKAGES="profiles,ads-api-v1-all"

For more information, see Amazon's Campaign Management Overview.

More than an MCP server

Everything above is the tool layer: the Amazon Ads operations an agent can call. Kuudo adds three layers around it, and a comparison that only counts operations misses them.

Kuudo provides higher-level workflows above the MCP tool layer through Skills. A Skill composes these operations into a repeatable Amazon procedure, grounds it with Atlas, and waits for a human before it changes anything.

Installation

We recommend running Amazon Ads API MCP with 🐳 Docker. Build the image from this repository (there is no supported docker pull for a third-party registry image here—use the Dockerfile and docker-compose.yaml in the repo).

For bespoke installs such as pip, source checkouts, custom virtualenvs, non-Docker deployments, or Code Mode extra handling, use the Installation Guide. The README keeps the Docker quick path short; INSTALL.md has the full install matrix and verification steps.

git clone https://github.com/KuudoAI/amazon-ads-mcp.git
cd amazon-ads-mcp

Copy the environment template:

cp .env.example .env

Edit .env with your settings (including PORT if you change the default).

Start the server with Docker Compose (builds and tags amazon-ads-mcp:latest per docker-compose.yaml):

docker compose up -d

The server listens on the host port mapped from PORT in .env (.env.example uses 9080).

Check logs:

docker compose logs -f

Stop the server:

docker compose down

For verification, upgrading, developer setup, and non-Docker installation paths, see the Installation Guide.

Configuration

Operators / self-hosters: follow the steps below and your .env file. Chat users mostly interact via prompts once the server is connected; skim Advertiser Profiles & Regions for profile and region behavior.

Amazon Ads requires that all calls to the API are authorized. If you are not sure what this means, you should read the Amazon docs:

• Amazon Ads API onboarding overview: https://advertising.amazon.com/API/docs/en-us/guides/onboarding/overview
• Getting started with the Amazon Ads API: https://advertising.amazon.com/API/docs/en-us/guides/get-started/overview

There are three paths for connecting to the API;

1. Bring Your Own App (BYOA)
2. Use OpenBridge-managed authorizations
3. Use Kuudo-managed authorizations

Bring Your Own Amazon Ads API App

If you have your own Amazon Ads API app, or want to create one, the process is detailed below.

1. Register Your Application with Amazon

1. Go to the Amazon Developer Console
2. Create or select your Login with Amazon application
3. Note your Client ID and Client Secret
4. Set your callback URL to "Allowed Return URLs". This must match the host and port of this MCP server’s HTTP listener (see PORT in .env; .env.example uses 9080):
   • For production: https://your-server.com/auth/callback
   • For local development: http://localhost:<PORT>/auth/callback (e.g. http://localhost:9080/auth/callback)

Once you have your app secured and approved by Amazon, you will need the client ID and secret:

# Amazon Ads API Credentials (required)
AMAZON_AD_API_CLIENT_ID="your-client-id"
AMAZON_AD_API_CLIENT_SECRET="your-client-secret"

Make sure these are in your .env file. Also, make sure you set your authorization method to direct in the same .env:

AUTH_METHOD=direct

Complete OAuth Flow

To authorize your connection to Amazon, you need to complete an OAuth workflow as an end user. First, you need to set your region. Authorization occurs at the region level and not setting your region may cause a failure. The server will default to the na region. You can manually set the region with tool set_active_region.

• Tool: set_active_region
• Parameters: na | eu | fe

Example prompt: "Set my current region to eu"

Step 1: Start OAuth

To connect to Amazon Ads API, you use an MCP tool to start your OAuth flow

• Tool: start_oauth_flow
• Example prompt: "Start my OAuth flow"

Step 2: Redirect to Amazon Ads

In this example, you are prompted to click the link that will open a browser window and request approval at Amazon.

Step 3: Approve Request

In the browser window, Amazon will prompt that you approve the request to connect.

Step 4: Success

If all goes well, you will see a success response. You can close the browser window and go back to your client. If you see something else, attempt the process again and confirm all your configuration elements are correct

Step 5: Confirmation

To confirm that your MCP server is connected to the Amazon Ads API, check your OAuth status

• Tool: check_oauth_status
• Example prompt: "Check my OAuth status"

You are ready to start interacting with the Amazon Ads API system!

Partner Applications: Token Authentication

You can configure your client, like Claude, to use authentication by supplying a valid access token. This is most appropriate for service accounts, long-lived API keys, CI/CD, applications where authentication is managed separately, or other non-interactive authentication methods.

Openbridge Partner App

As an Ads API Partner application provider, Openbridge offers a ready-to-go gateway to the Amazon Ads API. You log into your Openbridge account, provision a token, then set your token in your client config (see below).

First, set Openbridge as the auth method:

AUTH_METHOD=openbridge

That is it for the server config. To access the server, you need configure the client, like Claude Desktop, to pass the token directly. (see Example MCP Client: Connect Claude Desktop)

Authorized Amazon Accounts

Your Amazon authorizations reside in Openbridge. Your first step in your client is to request your current identities: "List my remote identities". Next, you would tell the MCP server to use one of these identities: "Set my remote identity to <>". You can then ask the MCP to List all of my Amazon Ad profiles linked to that account. If you do not see an advertiser listed, set a different identity.

Kuudo Partner App

Kuudo can vend short-lived Amazon Ads credentials from organization-managed Amazon connections. Configure the server with a Kuudo M2M API key:

AUTH_METHOD=kuudo
KUUDO_API_BASE_URL=https://amazon-spapi-dev.kuudo.ai/
KUUDO_API_KEY=sk_xxxxxx
KUUDO_PROVIDER=amazon_ads

The API key requires the amazon-connections:read and amazon-tokens:vend scopes. Set KUUDO_REMOTE_IDENTITY_ID to select a default connection at startup, or use the identity tools to list and select a connection before calling Amazon Ads operations.

Public HTTP deployments must also configure MCP_INBOUND_TOKEN or trusted proxy HMAC authorization for MCP callers.

Set Your Amazon Ads MCP Packages

To activate specific packages, set a comma-separated package list. For a streamlined default with unified Ads API v1 coverage:

• AMAZON_AD_API_PACKAGES="profiles,accounts-ads-accounts,reporting-version-3,amc-workflow,ads-api-v1-all"

Here is the list of tool packages available in the server:

• profiles
• test-account
• forecasts
• brand-stores-management
• campaign-manage
• accounts-manager-accounts
• accounts-ads-accounts
• accounts-portfolios
• accounts-billing
• accounts-account-budgets
• audiences-discovery
• reporting-version-3
• brand-benchmarks
• brand-metrics
• stores-analytics
• sponsored-products
• sp-suggested-keywords
• sponsored-brands-v4
• sponsored-brands-v3
• sponsored-display
• dsp-measurement
• dsp-advertisers
• dsp-audiences
• dsp-conversions
• dsp-target-kpi-recommendations
• amazon-attribution
• audience-insights
• partner-opportunities
• tactical-recommendations
• persona-builder
• creative-assets
• change-history
• data-provider-data
• products-metadata
• products-eligibility
• unified-pre-moderation-results
• moderation-results
• amazon-marketing-stream
• locations
• exports-snapshots
• marketing-mix-modeling
• reach-forecasting
• amc-administration
• amc-workflow
• amc-rule-audience
• amc-ad-audience
• ads-api-v1-sp (Beta)
• ads-api-v1-sb (Beta)
• ads-api-v1-dsp (Beta)
• ads-api-v1-sd (Beta)
• ads-api-v1-st (Beta)
• ads-api-v1-all
• ads-api-v1-beta (Beta)

You will note that some are broken up into smaller groupings. For example, Amazon Marketing Cloud has bundles; amc-ad-audience, amc-administration, amc-rule-audience, and amc-workflow. This is done to create efficiencies and optimizations that reduce context limits in many AI clients.

Understanding Amazon Ads MCP Tools

Amazon Ads MCP tools have prefixes (like cm_ for Campaign Management or amc_ for Amazon Marketing Cloud) to help organize the specific Ads API operation.

Example prefixes:

• cm_ → campaign/advertising APIs
• amc_ → AMC-related APIs
• dsp_ → DSP APIs
• sd_ → Sponsored Display
• ams_ → Amazon Marketing Stream
• spv1_ → Sponsored Products v1 (Beta)
• sbv1_ → Sponsored Brands v1 (Beta)
• dspv1_ → Amazon DSP v1 (Beta)
• sdv1_ → Sponsored Display v1 (Beta)
• stv1_ → Sponsored Television v1 (Beta)

This will translate into collections of tools that align with the API operations that are available:

Campaign Management (cm_)

• cm_QueryCampaign — Query campaigns
• cm_CreateCampaign — Create campaign
• cm_UpdateCampaign — Update campaign
• cm_DeleteCampaign — Delete campaign
• cm_QueryAdGroup — Query ad groups

Sponsored Products (sp_)

• sp_listProductAds — List product ads
• sp_createKeywords — Create keywords
• sp_updateBids — Update keyword bids
• sp_getNegativeKeywords — Get negative keywords

AMC Workflows (amc_)

• amc_listWorkflows — List AMC workflows
• amc_executeWorkflow — Run workflow
• amc_getWorkflowStatus — Check workflow status

Users would see tools like:

• "List my Amazon Ads campaigns"
  → Claude uses: cm_QueryCampaign

• "Create an AMC workflow"
  → Claude uses tools such as amc_executeWorkflow (after discovery; exact names depend on your enabled packages)

• "Export my sponsored products ads data" → Claude uses: export_createAdExport

📥 Downloading Reports & Exports

When tools fetch a report or export from Amazon, they download the file onto the machine running the MCP server (local disk on the server host—Docker volume, VM, or bare metal—not your laptop unless the server runs there). The default base directory is ./data (override with AMAZON_ADS_DOWNLOAD_DIR). Files are organized under data/profiles/{profile_id}/….

MCP tools list those server-side files and build URLs; clients then pull the bytes over HTTP (GET /downloads, GET /downloads/{relative-path}). Nothing streams straight from Amazon into the MCP client without hitting server storage first in these flows.

Use HTTP transport for downloads: run with --transport http (or Docker on port 9080). stdio mode does not register /downloads, and the get_download_url tool returns an error without HTTP.

Reports vs exports

• V3 async reports: Use the OpenAPI tools createAsyncReport (POST) to create, getAsyncReport (GET) to poll status, then download_export to save the completed file on the server into profile storage.
• Exports (e.g. ad exports): Create the job with the OpenAPI tools (e.g. export_CampaignExport), poll with export_GetExport, then call download_export with the completion URL so the file is fetched from Amazon and written on the server (same storage layout as reports).

Workflow

1. Set active profile (set_active_profile) so storage and HTTP access are scoped.
2. Create the job: Use the appropriate OpenAPI tool to create a report or export.
3. Poll for completion: Use the corresponding GET tool to check status until COMPLETED.
4. Download: Call download_export with the download URL from the completed response.
5. Discover or link: MCP tools list_downloads and get_download_url.
6. Fetch bytes: HTTP GET on the URL from step 5 (browser, curl, etc.).

Subpaths under the profile folder depend on report/export type (for example reports/…, exports/…). URLs look like http://localhost:9080/downloads/{relative-path} where {relative-path} matches list_downloads (illustrative example: reports/async/some-report.json.gz).

Direct HTTP (curl): /downloads uses the active profile in the running server (the one set via MCP in that process). If you see No active profile, call set_active_profile through MCP against the same server first. When no auth manager is configured, a numeric ?profile_id= query parameter may be accepted instead—see your deployment settings.

Example Prompts

HTTP Download API

# List available downloads (paths and URLs are profile-scoped)
curl http://localhost:9080/downloads

# Optional filter
curl 'http://localhost:9080/downloads?type=reports'

# Download one file (example path — use paths from list_downloads)
curl -O 'http://localhost:9080/downloads/reports/async/example-report.json.gz'

# If AMAZON_ADS_DOWNLOAD_AUTH_TOKEN is set
curl -H "Authorization: Bearer your-token" -O 'http://localhost:9080/downloads/exports/campaigns/example-export.json'

Profile Isolation

• Files live under data/profiles/{profile_id}/.
• Download tools and /downloads only see the active profile’s files.
• Set the profile first: "Set my active profile to 123456789".
• Legacy files outside profiles/ are not listed when using profile-scoped mode; move them under data/profiles/{profile_id}/ for access.

Advertiser Profiles & Regions

Setting Your Advertiser Profile

Per Amazon: Profiles play a crucial role in the Amazon Ads API by determining the management scope for a given call. A profile ID is a required credential to access an advertiser's data and services in a specific marketplace.

You may not know what profile(s) authorization grants you access to. You can list all advertising profiles accessible by your authorization:

• Tool: ac_listProfiles
• Example prompt: "List my advertiser profile ids"

Warning: Large accounts can return very large profile lists that may exceed client context limits. Prefer these bounded tools for discovery:

• Tool: summarize_profiles — "Summarize my advertiser profiles"
• Tool: search_profiles — "Find profiles with Acme in the name in US"
• Tool: page_profiles — "Show the first 20 UK profiles"
• Tool: refresh_profiles_cache — "Refresh my profile list cache"

Response includes profile details:

• profileId, countryCode, currencyCode
• dailyBudget, timezone
• accountInfo (type: seller/vendor/agency)

Let's assume your list included profile ID 1043817530956285. You can check for more details by getting profile details to confirm this is the one you want to use.

• Tool: ac_getProfile
• Example prompt: "Get the details for my profile_id: 1043817530956285"

Assuming this is the profile you want to use, you need to set the profile Amazon requires for API calls:

• Tool: set_active_profile
• Example prompt: "Set my active profile id to 1043817530956285"

When you set the profile, it determines:

• Which account's data you access
• Currency and timezone for reports
• Available campaigns/ads/keywords

The profile ID will be set in the background for the duration of your session. Repeat the process if you want to switch to a new profile.

Most calls to the Amazon Ads API require a Region. Each advertiser profile ID is associated with an advertising account in a specific region/marketplace.

The region is part of an advertiser profile. When you set an advertiser profile with set_active_profile, it will set the region that is associated with the profile automatically.

• Tool: set_active_profile

Example prompt: "Set my active advertiser profile to 111111111111"

Since profile ID 111111111111 is based in na, the region will be set based on the profile region.

Set Active Region

The Amazon Ads MCP server includes tools for managing API regions as defaults and dynamically, allowing you to switch between North America (na), Europe (eu), and Far East (fe) regions without restarting the server.

When you set a region, the system automatically:

1. Updates API endpoints - Routes API calls to the correct regional endpoint
2. Updates OAuth endpoints - Uses the correct token refresh endpoint for the region
3. Clears cached tokens - Ensures fresh authentication for the new region
4. Preserves other settings - Keeps profile ID and identity settings intact

IMPORTANT: Avoid Region Mismatch: If you attempt to set a region that is not associated with your advertiser profile, the Ads API will reject your requests. For example, if a profile ID is attached to na and you manually set the region to eu, you have created a mismatch which will cause API request failures.

Get Active Region

If you are not sure what region is set, you can check for the region

• Tool: get_active_region
• Returns: Current region, endpoints, and configuration source

Example prompt: "What is my current active region?"

Example MCP Client: Connect Claude Desktop

Navigate to Connector Settings

Open Claude in your browser and navigate to the settings page. You can access this by clicking on your profile icon and selecting “Settings” from the dropdown menu. Once in settings, locate and click on the “Connectors” section in the sidebar. This will display your currently configured connectors and provide options to add new ones.

Edit your Claude Desktop configuration file:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%\Claude\claude_desktop_config.json
Linux: ~/.config/Claude/claude_desktop_config.json

In this example, we show how to use the bearer token using the Openbridge API key. Add this configuration to your mcpServers section:

{
  "mcpServers": {
    "amazon_ads_mcp": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote@latest",
        "http://${HOSTNAME}:${PORT}/mcp/",
        "--allow-http",
        "--header",
        "Authorization:Bearer ${OPENBRIDGE_API_KEY}",
        "--header",
        "Accept:application/json,text/event-stream",
        "--debug"
      ],
      "env": {
        "HOSTNAME": "your_hostname",
        "PORT": "your_server_port",
        "MCP_TIMEOUT": "120000",
        "MCP_REQUEST_TIMEOUT": "60000",
        "MCP_CONNECTION_TIMEOUT": "10000",
        "MCP_SERVER_REQUEST_TIMEOUT": "60000",
        "MCP_TOOL_TIMEOUT": "120000",
        "MCP_REQUEST_WARNING_THRESHOLD": "10000",
        "OPENBRIDGE_API_KEY": "your_openbridge_token_here"
      }
    }
  }
}

Note: Replace hostname, port and your_openbridge_token_here with your actual OpenBridge token.

IMPORTANT: Cursor and Claude Desktop (Windows) have a bug where spaces inside args aren't escaped when it invokes npx, which ends up mangling these values. You can work around it using: mcp-remote custom headers documentation.

The config would look something like this:

{
  "mcpServers": {
    "amazon_ads_mcp": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "http://${HOSTNAME}:${PORT}/mcp/",
        "--allow-http",
        "--header",
        "Authorization:${AUTH_HEADER}",
        "--header",
        "Accept: application/json, text/event-stream"
      ],
      "env": {
        "HOSTNAME": "your_hostname",
        "PORT": "your_server_port",
        "MCP_TIMEOUT": "120000",
        "MCP_REQUEST_TIMEOUT": "60000",
        "MCP_CONNECTION_TIMEOUT": "10000",
        "MCP_SERVER_REQUEST_TIMEOUT": "60000",
        "MCP_TOOL_TIMEOUT": "120000",
        "MCP_REQUEST_WARNING_THRESHOLD": "10000",
        "AUTH_HEADER": "Bearer <your_openbridge_token_here>"
      }
    }
  }
}

Here is another example, which can be used if you are using OAuth since the OPENBRIDGE_API_KEY is not needed:

{
  "mcpServers": {
    "amazon_ads_mcp": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote@latest",
        "http://localhost:9080/mcp/",
        "--allow-http"
      ],
      "env": {
        "MCP_TIMEOUT": "120000",
        "MCP_REQUEST_TIMEOUT": "60000",
        "MCP_CONNECTION_TIMEOUT": "10000",
        "MCP_SERVER_REQUEST_TIMEOUT": "60000",
        "MCP_TOOL_TIMEOUT": "120000",
        "MCP_REQUEST_WARNING_THRESHOLD": "10000"
      }
    }
  }
}

Note: For various Claude configurations similar to what was shown above, see the MCP Remote docs for the latest settings/options.

Restart Claude Desktop

After saving the configuration file, restart Claude Desktop to load the new MCP server.

⚠️ Context Limits and Active MCP Server Tools

MCP tool registration and use can impact your AI systems usage limits. Usage limits control how much you can interact with an AI system, like Claude, over a specific time period. As Anthropic states, think of the amount of information/data used as drawing down on a "conversation budget". That budget determines how many messages you can send to your AI client, or how long you can work, before needing to wait for your limit to reset.

MCP Server tools contribute metadata like titles, descriptions, hints, and schemas to the model's context. This metadata is loaded into the LLM’s context window, which acts as its short-term working memory.

Each client, like Claude, has a fixed-size context window. This defines the maximum amount of information it can process in a single interaction—including user prompts, system instructions, tool metadata, and any prior messages.

The more tools you activate, the more of that limited space gets consumed up front. When you activate many tools, their combined schema and config payloads can significantly use up this context and you may quickly hit the context ceiling. This is when you’ll start seeing errors or warnings about exceeding the chat length limit.

The Amazon Ads MCP provides coverage across the entire API. As a result, there can be 100s of tools!

• More tools = less room for user interaction: Activating unnecessary tools reduces available space for your actual prompt or data.
• Start small: Activate only what you need for the current task. You can always add more later.

If you're encountering unexpected length issues, review which tools are active. Trimming unused ones can help minimize context use.

Code Mode

Code Mode is a feature that dramatically reduces tool token consumption. Instead of loading every tool schema into the LLM's context upfront, Code Mode replaces the full catalog with four lightweight meta-tools. The LLM discovers tools on demand and executes them via sandboxed Python.

Token Impact

Illustrative comparison for a representative default-sized catalog; actual tool counts and token use depend on AMAZON_AD_API_PACKAGES, server version, and tokenizer.

How It Works

Code Mode uses a 4-stage discovery pattern:

Stage 1: tags          → Browse categories: "campaign-management (20), accounts (14), ..."
Stage 2: search        → Find tools: search("campaigns") → cm_listCampaigns, cm_createCampaign, ...
Stage 3: get_schema    → Get parameters: get_schema("cm_listCampaigns") → full input schema
Stage 4: execute       → Run code: sandboxed Python with await call_tool("cm_listCampaigns", {...})

All 200+ tools remain fully accessible. The LLM simply discovers and calls them on demand rather than loading every schema upfront.

Activating Code Mode

Code Mode defaults to CODE_MODE=true. To use the full tool catalog in context instead, set:

CODE_MODE=false

Docker Compose — add to your environment section:

environment:
  - CODE_MODE=false

Docker Run (map the host port to the container PORT, e.g. 9080):

docker run -d --env-file .env -e CODE_MODE=false -e PORT=9080 -p 9080:9080 amazon-ads-mcp:latest

Local Development (entry point is python -m amazon_ads_mcp.server, which runs the MCP server module):

CODE_MODE=false uv run python -m amazon_ads_mcp.server --transport http --port 9080

Configuration Options

Using Code Mode

Once enabled, your MCP client (Claude Desktop, Claude Code, etc.) will see only four tools instead of the full catalog.

Browse available categories:

"What tool categories are available?"

The LLM calls tags and sees categories like campaign-management, sponsored-products, programmatic-dsp, accounts, reporting, etc.

Find specific tools:

"Find tools for managing campaigns"

The LLM calls search("campaigns") and gets matching tool names with brief descriptions.

Get tool details:

"Show me the parameters for creating a campaign"

The LLM calls get_schema("cm_CreateCampaign") and gets the full input schema.

Execute operations:

"List my enabled campaigns"

The LLM calls execute with Python code:

result = await call_tool("cm_QueryCampaign", {
    "body": {"stateFilter": {"include": ["ENABLED"]}}
})

Tool Categories

When the LLM calls tags, it sees human-readable categories mapped from the API prefixes:

When to Use Code Mode

Use Code Mode when:

• You have many tool packages activated and are hitting context limits
• You want maximum context space for conversation and data analysis
• Your MCP client supports tool discovery patterns (Claude Desktop, Claude Code)

Use Standard Mode when:

• You have a small number of packages activated (1-3)
• You need the fastest possible tool invocation (no discovery step)
• Your workflow uses the same few tools repeatedly

Verifying Code Mode

After starting the server with CODE_MODE=true, verify it is active:

# Check server logs for confirmation
docker logs <container> 2>&1 | grep "Code mode"
# Expected: "Code mode active: 4 meta-tools exposed (tags, search, get_schema, execute)"

In your MCP client, you should see only four tools: tags, search, get_schema, and execute. If you see the full tool catalog, code mode is not active — check your environment variable.

Tool Audit

Use tool_audit to measure token footprint for tool catalogs and code mode behavior.

Baseline Audit

.venv/bin/python -m amazon_ads_mcp.tool_audit \
  --url http://127.0.0.1:9080/mcp \
  --limit 20

JSON output:

.venv/bin/python -m amazon_ads_mcp.tool_audit \
  --url http://127.0.0.1:9080/mcp \
  --format json \
  --limit 0 \
  > /tmp/tool_audit.json

Code Mode Probe Audit

When code mode is active, use probe mode to estimate on-demand schema fetch cost:

.venv/bin/python -m amazon_ads_mcp.tool_audit \
  --url http://127.0.0.1:9080/mcp \
  --format json \
  --limit 0 \
  --code-mode-probe \
  --probe-limit 50 \
  --probe-queries "campaign,ad,report,profile,dsp,amc,brand,targeting" \
  > /tmp/tool_audit_code_mode.json

Quick Comparison

jq '{tool_count,total_tool_tokens,context_window_percent}' /tmp/tool_audit.json
jq '{tool_count,total_tool_tokens,context_window_percent,code_mode_probe}' /tmp/tool_audit_code_mode.json

Key fields:

• total_tool_tokens: Upfront token load from tools/list
• context_window_percent: Share of configured context window
• code_mode_probe.total_schema_tokens: Total sampled schema-fetch token cost in code mode
• code_mode_probe.avg_schema_tokens: Average token cost per fetched schema

Background Tasks

Many Amazon Ads API operations are long-running — report generation, export creation, AMC workflow execution, audience processing, and DSP measurement studies can take seconds to minutes. Background tasks allow these operations to run without blocking the conversation.

How It Works

Background tasks are enabled by default. When a client requests background execution for any tool:

1. Start: The tool returns immediately with a task ID
2. Track: The client polls for progress updates at server-suggested intervals
3. Retrieve: The client fetches the result when the task completes

All async tools in the server support background execution automatically. The client decides per-call whether to run a tool in the foreground (wait for result) or background (get task ID, poll later).

Built-in workflow tools

These tools orchestrate multi-step workflows (download and HTTP access to server-side files):

These tools include progress reporting so clients can track each stage of the workflow.

Configuration

To disable background tasks:

ENABLE_TASKS=false

Docker Compose — add to your environment section:

environment:
  - ENABLE_TASKS=false

Backend

The server uses an in-memory task backend by default. Tasks are tracked for the lifetime of the server process. If the server restarts, in-progress tasks are lost.

For persistent task tracking across restarts, configure a Redis backend:

FASTMCP_DOCKET_URL=redis://localhost:6379

Example Prompts

Frequently asked

Does Kuudo only expose Amazon Ads API operations?

No. This server exposes the Amazon Ads operations themselves. Kuudo Skills package them into higher-level workflows: reusable, versioned, multi-step procedures that can combine tools from several MCP servers, retrieve Amazon and account knowledge through Agent Atlas, use Agent Flow for durable data work in your own cloud, and wait for human approval before anything goes live. Skills run from any MCP client, from code, or from a CLI.

Do I have to use a prebuilt Skill?

No. An agent can compose these operations directly for one-off work. Skills exist for the procedures you run repeatedly and want to run the same way every time. Pick one, watch it run, then fork it for your account, or write your own against these tools: every Skill is a plain open-format file, not a proprietary one.

Is this the whole product?

No. This repository is the Amazon Ads tool layer. Kuudo also runs MCP servers for Selling Partner and Vendor Central, the Skills workflow layer, Agent Atlas for knowledge, and Agent Flow for durable data work. See how the pieces fit.

Troubleshooting

Server not connecting?

• Ensure the Docker container is running: docker compose ps
• Check server logs: docker compose logs -f
• Verify the host port matches PORT in .env (.env.example uses 9080; your mapping must match)

Authentication errors?

• Check your OpenBridge token is valid
• Ensure the token is properly set in the environment
• Verify your Amazon Ads API access

Claude not recognizing the server?

• Restart Claude Desktop after configuration changes
• "Reload Page" in Claude Desktop if the MCP is not active
• Check the JSON syntax is valid
• Ensure the server name matches exactly

Documentation map

📄 License

This project is licensed under the MIT License — see the LICENSE file for details.

Security: report vulnerabilities through GitHub private security advisories for this repository.