Commerce Gate MCP
Commerce Gate MCP evaluates commerce proposals in Shadow Mode and retrieves decision evidence. Its local STDIO profile also provides an ERP policy and agent-budget guard. When enabled, the separate hosted OAuth reviewer offers seven authenticated tools for evaluation, policy readiness and evidence, using only decision:read, decision:run and policy:read. Neither profile executes marketplace or ERP writes. Stop on HOLD, BLOCK, ESCALATE or errors.
Decionis CommerceAgent
Decionis CommerceAgent is the name for an AI agent that drafts commerce decisions through Commerce Gate. It watches your listings, orders and refunds, proposes the next action with the facts behind it, asks Commerce Gate whether that action clears policy, and stops on HOLD, ESCALATE or BLOCK. A held decision reaches a person on your desk with CommerceAgent's draft beside it. Your repricer, ERP, marketplace connector or that person performs the action, and every answer is recorded.
Today CommerceAgent runs inside your agent host with the published Commerce Gate MCP below. It pairs with Decionis PolicyAgent, which authors the policy CommerceAgent operates under. It is not a shopping assistant for buyers, and it is not a payment gateway: it never processes payments, holds funds or moves money.
Read listings, orders, refunds and what Shadow Mode held recently.
State one exact action with its facts: SKU, prices, quantities, cost, balances.
Submit it to Commerce Gate and read the outcome and the evidence reference.
Stop on HOLD, ESCALATE, BLOCK, an error or an ambiguous answer.
On PROCEED, your own tools or a named person perform the action. Keep the dossier reference with the record.
Marketplace operations, pricing, fulfillment, finance, and automation teams get one set of commerce evaluation outcomes and a separate binary ERP guard without giving the MCP direct execution power.
Ask whether a price change, stock change, order acceptance, fulfillment step, promotion change, refund or return clears policy. Commerce Gate MCP submits evaluations in Shadow Mode; decision evidence is returned where supported.
Capability discovery maps Walmart Commerce Evaluation, Price Guard, Order Guard, order interception, margin protection, release, hold, and cancel to bounded Shadow preflights. The MCP does not inspect a merchant connection or execute the connected Walmart path.
Retrieve a Decision Dossier or proof packet by identifier without reaching into a production database or marketplace account.
List and summarize recent advisory evaluations so an operator can understand patterns before enabling enforcement elsewhere.
Send one complete bounded D365 transaction to the enforced binary policy and agent-budget guard without posting, releasing, or modifying it in the ERP.
Check the bounded SaaS offer, plan, public URL, marketplace identity, preview or test, and digest-backed release-evidence packet before a person submits it. Microsoft requires Entra identity; AWS requires a Login with Amazon identity linked to the seller AWS account and IAM role. The MCP never calls a marketplace.
It never writes to a marketplace or ERP. Your integration must honor the answer and obtain authority for the final action.
MCP separates host, client and server responsibilities. A server's exposed tools determine what it can do; protocol support alone does not grant execution permission.
Plain questions map to seven published Shadow action contracts, the enforced ERP guard, and a Microsoft Partner Center and AWS Marketplace SaaS offer-submission preflight. The agent evaluates, validates, and reads evidence; a person or the connected platform still executes.
“Would repricing DC-AFR-55 to $89 on Walmart still clear our margin floor after the referral fee?”
The agent submits the proposed price and supplied cost facts in Shadow Mode, then receives a policy outcome and evidence reference. The response does not publish the price or call Walmart.
commercegate_evaluate_action“Can the Walmart Order Guard intercept this order, hold it for margin protection, or release or cancel it?”
The agent evaluates the exact order or fulfillment proposal in Shadow Mode. Order interception and margin protection map to ORDER_ACCEPTANCE; release maps to FULFILLMENT_ACTION with acknowledge, while hold and cancel use their matching fulfillment actions. The result never calls Walmart.
commercegate_evaluate_action“Would accepting this $4,000 order after a $300 discount clear policy with $2,900 in estimated cost?”
The agent submits the supplied gross amount, discount, and estimated cost in Shadow Mode. HOLD means a person reviews the economics before any separate execution step.
commercegate_evaluate_action“Would this stock change leave only seven units and create oversell risk?”
The agent sends the current and new quantity and gets a verdict against your inventory floor. Whether the platform can then stop the write depends on that platform; the verdict is recorded either way.
commercegate_evaluate_action“Should this $2,850 refund request be held for review?”
The agent submits the refund request without issuing it. Any later refund requires a separately connected marketplace path with the required API, scope, lifecycle checks, and explicit authority.
commercegate_evaluate_action“What would Shadow Mode have held this month, and why?”
The agent lists and summarizes recent Shadow evaluations: outcome counts and rates, near misses, reason breakdowns, and candidate enforcement paths.
commercegate_summarize_shadow_reports“Finance wants the signed evidence for dossier 8d4d…e201.”
The agent fetches the Decision Dossier and proof packet by ID. Verification is evidence, never permission to execute.
commercegate_get_proof_packet“Does this complete Business Central sales document clear policy and the agent budget?”
The agent calls the enforced ERP guard with the bounded transaction and ERP region. ALLOW or BLOCK applies to that exact request; the MCP does not release or post the document.
commercegate_validate_erp_transaction“Is this Commerce Gate SaaS offer packet ready for manual Microsoft Partner Center or AWS Marketplace submission?”
The agent checks the exact offer, plans, markets, public URLs, platform identity, preview or test confirmation, and release-evidence digests in Shadow Mode. It does not create, change, upload, or submit an offer in a marketplace.
commercegate_evaluate_marketplace_offer_submissionAgent operating rule
Confirm the supported action shape and configuration readiness.
Submit one exact proposed commerce action, a complete D365 transaction when the ERP guard is the intended authority, or the complete marketplace SaaS offer packet before manual submission.
Stop on HOLD, ESCALATE, BLOCK, errors, or ambiguity and surface the evidence.
Tool catalog
commercegate_describe_capabilitiescommercegate_validate_erp_transactioncommercegate_evaluate_actioncommercegate_get_dossiercommercegate_get_proof_packetcommercegate_list_shadow_reportscommercegate_summarize_shadow_reportscommercegate_evaluate_marketplace_offer_submissionPublic npm transport
Install and run the pinned @decionis/commerce@0.2.0 release with npx -y @decionis/commerce@0.2.0. The package runs as a local STDIO process. Calls use configured DECIONIS_API_KEY and DECIONIS_ORG_ID, or access from the private AgentOps store. Environment configuration takes precedence. The API origin defaults to https://api.decionis.com.
On macOS and Linux, the first valid local Shadow evaluation can create provisional access with trial limits. Discovery, evidence reads and ERP calls never create access; ERP authorization requires owned access. Windows requires existing credentials. Set AGENTOPS_AUTO_PROVISION=0 to disable new provisioning while retaining existing access.
The npm package is a local adapter over the Decionis decision and evidence APIs. Hosted OAuth connections use the separate reviewer profile below. The broader hosted Protocol MCP remains a separate parent interface.
A verdict from the MCP says whether an action clears policy. Whether the platform can then enforce it depends on that platform: the marketplace has to offer the API, you have to grant the scope, and the connector has to support the action. Each platform page states what is enforced today.
View the package on npm Decionis Protocol MCPHosted OAuth reviewer
Hosted OAuth reviewer: POST-only Streamable HTTP at https://api.decionis.com/v1/commerce/mcp. Accounts sign-in and workspace consent grant only decision:read, decision:run, policy:read. Its seven authenticated tools evaluate in SHADOW, read policy readiness and retrieve evidence. This profile has no ERP transaction guard, agent-budget authorization, marketplace offer-submission tool or external writes. Check live resource and authorization-server metadata before connecting; public clients use S256 PKCE and dynamic registration only when advertised.
commercegate_describe_capabilitiescommercegate_evaluate_actioncommercegate_get_dossiercommercegate_get_proof_packetcommercegate_list_shadow_reportscommercegate_summarize_shadow_reportscommercegate_get_policy_readinessEligible direct and AWS Marketplace Commerce subscriptions can connect through Accounts. Microsoft entitlements use their existing purchase path.
Membership and subscription access are checked on every token use. A connection does not establish public directory approval or permission to execute a commerce action.
Start with Shadow Mode, inspect the dossier, and keep marketplace execution outside the MCP boundary.
It evaluates price changes, stock changes, order acceptance, fulfillment steps, promotion changes, refund requests and return authorizations. Price and order inputs have native margin mappings; the other five use Protocol policy checks that depend on your tenant policy and the signals sent. These commerce-action checks run in Shadow Mode.
Decionis CommerceAgent is the name for an AI agent that runs commerce operations through Commerce Gate. It observes listings, orders and refunds, proposes an action with its facts, asks Commerce Gate whether the action clears policy and stops on HOLD, ESCALATE or BLOCK. It runs in your agent host with the published MCP; your own tools or a named person perform the action. It is not a buyer-facing shopping assistant and not a payment gateway.
APPROVE maps to PROCEED, REJECT maps to BLOCK, and REVIEW or ESCALATE maps to HOLD. The ERP guard returns ALLOW or BLOCK. PROCEED and ALLOW are policy evidence, not user consent to execute. The MCP does not send the commercial action, and an error must be treated as HOLD.
Each platform has its own answer. Walmart order acknowledgment and Adobe Commerce checkout fail open: if Decionis can't be reached, the order or checkout goes through and the gap is recorded. Shopify checkout fails closed by default, rule by rule, so a margin rule keeps protecting you; you can set individual rules to fail open. Its post-order inventory check only records a decision and never touches the order. Business Central has one company-wide setting, FAIL_OPEN or FAIL_CLOSED, and defaults to FAIL_OPEN. When an AI agent asks the Commerce Gate MCP and gets an error instead of an answer, the agent must treat it as HOLD. Walmart refunds are rejected when the order is in the wrong state or the amount exceeds what is left to refund. A refund above the unattended limit in your policy escalates to a named approver, and a second refund on the same line is held for review; if the refund policy check itself cannot be reached, refunds up to the connector's fallback limit proceed and larger ones are held until it recovers.
Where your platform records a Decision Dossier, it explains the decision and links to signed evidence, and a proof packet helps you review it. Some observations, outage fallbacks and operator-confirmed actions have no dossier. A record that a check ran does not prove the price changed, the order shipped or the refund completed.