Atlas MCP
The Amazon knowledge
your agent didn't
have to earn.
Atlas indexes 7,589 documents across Ads, Seller, Vendor, and a hand-built decision rulebook. Your agent queries semantically — answers come back in milliseconds, with citations, in a structure the model can use.
What's in the corpus
7,589 documents covering every Amazon system your team works with.
Amazon Ads & AMC
Sponsored Ads, DSP, AMC audiences, attribution, SQL templates, conversion paths, audience segmentation.
Seller Central
FBA fulfillment, Brand Registry, listings, Buy Box, fees, inventory, shipping, returns, tax.
Vendor Central
Compliance, chargebacks, ARA/ABA reports, A+ content, Net PPM, EDI/API integrations, catalog operations.
Decision rulebook
IF/THEN bidding rules, sufficiency gates, conflict resolution, action enums — plus Amazon's category style guides for 11 retail verticals.
Why retrieval changes the agent
It looks up the answer before it answers.
Retrieves by meaning, not keyword
all-MiniLM-L6-v2 embeddings find the right playbook even when your question doesn't match the docs verbatim.
Filters before it answers
Every chunk carries domain, topic, doc type, source path, dataset references, and code language tags.
Citations on every claim
Section paths and code language are preserved at chunk time, so every recommendation links back to its source.
Rulebook, not improv
Scale ±5–10%, daily clamp ±20%, re-evaluate in 5 days. The agent consults the rulebook instead of guessing.
Compiles on first try
Knows conversions_for_audiences vs conversions_all, the right event_subtype codes, and the 14-day attribution wait.
Knows what's current
Long-term storage fees ended Jul 2024. Chargeback dispute window is 30 days. Sandbox doesn't populate SnS.
AMC SQL playbooks
Retrieved AMC SQL. Compiles on the first run.
AMC unlocks cross-channel attribution, audience overlap, and customer journey analysis. The SQL is the wall most teams stall against. Atlas indexes the playbooks Amazon publishes — and the operational details Amazon doesn't.
"Build me a DSP audience of people who added one of our hero ASINs to cart but didn't buy in the last 30 days."
Retrieves the cart-abandonment Audience IQ, knows it must use conversions_for_audiences (the only variant that allows SELECT user_id), and produces a query that compiles on the first try.
WITH asins (asin) AS (
VALUES ('B0XXXXXXX1'), ('B0XXXXXXX2')
),
purchase AS (
SELECT user_id, MAX(event_dt_utc) AS purchase_dt_max
FROM conversions_for_audiences
WHERE event_subtype = 'order'
AND tracked_item IN (SELECT asin FROM asins)
GROUP BY 1
),
atc AS (
SELECT user_id, MAX(event_dt_utc) AS atc_dt_max
FROM conversions_for_audiences
WHERE event_subtype = 'shoppingCart'
AND tracked_item IN (SELECT asin FROM asins)
GROUP BY 1
)
SELECT atc.user_id
FROM atc
LEFT JOIN purchase ON atc.user_id = purchase.user_id
WHERE atc_dt_max > purchase_dt_max
OR purchase_dt_max IS NULL;
"How much more do my Subscribe & Save customers spend compared to one-off buyers?"
Knows that Feb 2024 extended FSI to include repeatSnSOrder alongside snsSubscription and firstSnSOrder, and returns a working comparison query.
SELECT
CASE
WHEN event_subtype IN (
'snsSubscription','firstSnSOrder','repeatSnSOrder'
) THEN 'subscriber'
ELSE 'non-subscriber'
END AS user_type,
COUNT(DISTINCT user_id) AS users_that_purchased,
SUM(total_product_sales) AS total_sales,
SUM(total_product_sales) /
COUNT(DISTINCT user_id) AS average_spend
FROM conversions_all
GROUP BY 1;
"What's the actual journey our customers take across DSP, SP, SB, and SD before they convert?"
Pulls Customer Journey Analytics, identifies the three required tables (dsp_impressions, sponsored_ads_traffic, amazon_attributed_events_by_traffic_time), and outputs a result shaped for a Sankey diagram.
-- Path-to-Conversion grouped by campaign category
-- Joins DSP impressions, sponsored ads traffic,
-- and attributed events on user_id within window.
SELECT
path_signature,
campaign_category_path,
COUNT(DISTINCT user_id) AS converters,
SUM(total_product_sales) AS sales
FROM customer_journey_paths
GROUP BY 1, 2
ORDER BY converters DESC;
Beyond AMC
FBA fees. Chargebacks. Bid math. The unglamorous Amazon work, handled.
Vendor chargeback disputes
Pulls the exact 30-day dispute window, the second-dispute path, and the Smart Aggregation Processing 7-day preview cycle. Drafts the dispute in the evidence pattern Amazon's reviewers actually look for.
FBA inventory & storage
Knows long-term storage fees ended July 1, 2024. Quotes the active fees — monthly storage, aged-inventory surcharge, capacity limits in cubic feet — instead of the 2022 schedule.
Brand Registry conflicts
"Trademark already enrolled" → retrieves the specific resolution path: contact the existing Administrator via the application flow, or file via the FAQ's "use this form" mechanism if inactive.
Bid decisions, not bid vibes
Scale-winner fires only when ORD_g ≥ 10 AND ACOS_g ≤ 0.9 × TARGET_ACOS. Increase 5–10%, daily clamp ±20%, re-evaluate in 5 days. Conflict order: invalid → thin-data → safety → pull-back → scale.
Wired into a workflow
Retrieved and adapted, not generated.
Atlas is the retrieval layer beneath agents that already do work — pulling SP-API reports, querying AMC, drafting support replies, auditing campaigns, building audiences, planning inventory.
Your team asks in plain language
"Build me a high-value lookalike for Prime Day."
Agent searches the corpus semantically
Retrieves the AMC Lookalike Audiences for Promotional Events playbook, the high-value seed query template, and the 500–500k size constraints.
Agent grounds the response
Compilable seed query, named thresholds, the testing approach the playbook actually recommends, and the 48-hour DSP activation lag.
Your team runs it
Works the first time — because the answer wasn't generated, it was retrieved and adapted.
The bottom line
Atlas grounds the answer. Your agent ships it.
Every chunk traceable to its source. Every recommendation grounded in a citation. Every SQL template tested against the table names Amazon actually exposes.