Skip to content

Features

Four Amazon MCPs.
A Skills library. The full Kuudo.

The Ads and SP-API surfaces, plus Atlas for Amazon-specific retrieval and Crawl for live page context. Pre-built Skills on top. Invoke from a prompt, a script, or a scheduled run — all in your own cloud.

Quick start Read the docs
SP - auto
SB - brand
DSP - awareness
Ads MCP

Amazon Ads API MCP

One MCP for the full Amazon Ads API surface, built so agents stop guessing and start shipping.

Sponsored Products, Brands, Display, DSP, AMC, Attribution, Marketing Stream, and reporting behind one interface.

  • Self-correcting error envelopes
  • Token-conscious tool descriptions
  • Multi-region profile routing
Explore Ads MCP ->
orders.list
inventory.fba
listings.patch
fees.estimate
SP-API MCP

Amazon SP-API MCP

One MCP for the entire Selling Partner API surface, with marketplace-aware routing and corrective errors.

Catalog, Listings, Orders, Shipping, Reports, AWD, Finances, Messaging, and seller operations behind one interface.

  • Marketplace-aware routing
  • Corrective errors for agents
  • Broad SP-API package coverage
Explore SP-API MCP ->
Kuudo
doing
done
Atlas MCP

Amazon Agent Atlas

RAG for the Amazon corpus an agent actually needs to know.

Indexes your Amazon corpus — campaigns, listings, returns, fees, AMC queries, internal notes. The agent asks Atlas instead of scanning your warehouse from scratch.

  • Indexed Amazon corpus, queried in milliseconds
  • Tenant-scoped — your retrieval, your data, your cloud
  • Citations on every result — the agent shows its work
Explore Atlas MCP ->
Kuudo
doing
done
Crawl MCP

Amazon Agent Crawl

Live web context, fetched and formatted for agents.

Pulls a product page, a SERP, or a competitor's listing and returns clean, agent-ready content. No more training-set staleness when you need "now."

  • Fresh, not stale — fetched now, not from training data
  • Returns agent-ready content, not raw HTML
  • Works on any URL — Amazon PDPs, SERPs, competitors
Explore Crawl MCP ->
amc · audience
ad-ops sweep
sla rescue
bid pacing
asin sync
+ fork
Skills

Pre-built Skills

Verified, parameterized playbooks like AMC audience builds and ad-ops sweeps. Run from chat, no glue code.

Skills wrap multi-step automations into something an end user can actually run.

  • Verified happy paths with built-in guardrails
  • Approval queues for anything destructive
  • Fork, version, and ship your own
Explore skills ->
audience seed
path analysis
activation
privacy check · passed
AMC

Amazon Marketing Cloud

Run audience, attribution, incrementality, and signal-enrichment workflows — without writing a single SQL query.

Verified AMC Skills call your AMC instance, DSP seat, and partner signals through MCP.

  • Deep campaign findings and journeys
  • Bespoke DSP audience activation
  • Privacy-safe run logs and approvals
Explore AMC ->
"On Amazon-specific analytical tasks, Kuudo-grounded agents deliver 30–50% more accurate outputs than agents relying on training knowledge alone, and reduce expert review cycles by 2–4x."
Codex logo
Codex Solutions Architect Agentic

Why this Ads MCP is different

Production-ready on the first prompt.

Amazon's Ads API has more gotchas than almost any large API surface. This MCP catches each mistake and returns a corrective hint — wrong field name, deprecated body shape, wrong region. The agent's next call is right, not another guess. No three-week ramp before the model is useful.

×400 · invalid filter operator "eq"
↳use "EQUAL_TO" on /v1/sp/campaigns/list
retry · 200 · 14 campaigns
Self-correcting errors

The wrong move returns the right move

Every failure converts to a corrective hint. The model just reads it and moves on — no re-pasting, no thrashing.

campaignId→ campaign_idv3
adGroupId→ ad_group_idv3
state→ campaign_statev1
Field & alias resolver

Making sense of inconsistencies

Knows the same field can have three plausible names depending on the migration generation. Maps them at request time.

/v2/sp/campaignsdeprecated
/v1/sp/campaigns/listactive
body shape · POST · stateFilter→state
Deprecated-shape table

Ads deprecated? Here's the new one

Catches deprecated body shapes and filter operators before they hit Amazon, then suggests the active equivalent.

NAadvertising-api.amazon.com
EUadvertising-api-eu.amazon.com
FEadvertising-api-fe.amazon.com
Multi-region routing

NA, EU, APAC — automatic

Region endpoints resolved per profile. No hand-rolled host switching, no cross-region 401 spirals.

class Campaign(BaseModel):
campaign_id: str
daily_budget: Decimal
state: CampaignState
Type-safe envelopes

Agent-ready, not API soup.

Full type hints across campaigns, profiles, reporting, DSP, AMC workflows. The IDE catches it before the API does.

attempt 1
5 calls
attempt 2
2 calls
attempt 3
1 call
Compounding context

Every retry has direction

A five-tool-call debugging arc collapses into a two-call corrected one. The same mistake a week later doesn't re-burn context.

Context budget

Use 1% of the context window. Spend the rest on reasoning.

Compact tool descriptions, intelligent aliasing, and self-correcting envelopes mean the agent carries only what it needs into every turn. Pre-use cost drops from ~70,000 tokens to ~1,000. Context window usage stays under 1% — leaving the model 99% of its room for reasoning, retries, and longer tasks.

Pre-use cost comparison
Amazon hosted69,923
Ours1,035
98.5%
less pre-use tokens
Context window usage (200K window)
30.73%
Other MCPs
<1%
Ours

Compact tool descriptions, intelligent aliasing, and self-correcting envelopes mean the agent carries just what it needs, not the entire API schema, into every turn.

Every tool description, error envelope, and retry hint is compressed and structured to keep context lean. The result: an MCP that stays out of the way until it's needed, and costs less when it is.

You've felt this before

How to spot an agent that's burning context.

Vague errors trigger thrashing. Thrashing fills the context window. Useful state gets evicted. A one-call task balloons to 4 — sometimes 12 — round-trips before anything ships.

"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.

Quietly forgetting the fix

The agent loses the workaround it figured out three turns ago.

Five tool calls, four retries

None of them progress — just thrash on field names and shapes.

Re-pasting the same error

Into the chat, for the third time.

The agent is fine. The API surface is the one burning context — wrong field names, wrong body shape, wrong filter operator, wrong date placement, wrong advertiser-account format. The Ads MCP catches each one in the response envelope. The agent's next call is right, not another guess.

Private. Trusted. Owned.

An Amazon Verified Application
that runs in your cloud.

Reviewed under Amazon's Verified Applications program. Isolated per tenant in AWS, GCP, or Azure — your cloud, your IAM, your audit log. If you cancel Kuudo, the room you built keeps running.

aws gcp azure cf
vpc · peeredair-gap · off
Private

Runs in your cloud

Your Kuudo instance deploys into AWS, GCP, or Azure under your account. Per-tenant container, your VPC, your IAM. No shared compute, no cross-tenant access.

identity · verified
data-handling · reviewed
security · aws-audited
Trusted

Amazon Verified Application

Reviewed under Amazon's Verified Applications program — identity, data handling, and credential management audited against Amazon's bar. You inherit the partner posture instead of building it.

tenant-a
tenant-b
tenant-c
tenant-d
Owned

Yours to keep

Cancel Kuudo and the MCP servers, Skills, and artifacts you built stay in your cloud. No vendor lock-in. No "export your data" tickets.

Guides

How teams are running FBA audits, AMC builds, listing compliance, and fee estimates — without exporting a single spreadsheet.

Field notes from teams running Amazon work through Claude, ChatGPT, Cursor, n8n, and any MCP runtime — what's working, what to watch for, what to skip.

View all guides
/engineering

Counting New-to-Brand Customers in AMC Without Double-Counting the Same Buyer

How an Atlas-grounded AMC query counts genuine first-time buyers, why the conversion-time table is the only safe pick for a recurring workflow, and the five caveats the IQ docs scatter across four pages.

Thomas Spicer
8 min read
/engineering

Sponsored Ads x DSP Overlap: The 4-Way AMC Analysis

How an Atlas-grounded agent runs the improved Sponsored Ads and demand-side platform overlap query in AMC, why Sponsored Products exposure means impressions here, and the caveats that keep the result honest.

Thomas Spicer
8 min read
/engineering

Choosing the Right AMC Attribution Data Source for Custom Models

An operator asks for a custom attribution view in AMC. An agent grounded in Amazon Agent Atlas returns a five-row data-source matrix, six decision rules, and the footnotes that would have eaten the rest of the week.

Thomas Spicer
12 min read

Bring your Amazon knowledge into the game.

Connect ChatGPT. Pick the AMC and Sponsored Ads Skills. Stop re-explaining.
No credit card. No sales call.

Connect ChatGPT in 5 minutes