AI Agent · MCP

AI Agent and MCP: configure Marta in conversation

Connect Claude or ChatGPT to your loyalty program in one click. The Marta Agent speaks Italian with the operator and uses secure MCP tools to read dashboards, sales and promotion performance, draft campaigns and apply changes only after explicit confirmation.

The MCP server is native in Marta API: no opaque middleware, revocable tokens and a full audit. It is not MARTAIA: the Agent works on the platform from Claude, ChatGPT and other MCP clients.
Request a demo See the secure flow
Marta Agent with a draft waiting for explicit confirmation.
Visa Shopify Power BI OneSignal KPM Meta AWS

Native MCP Server

No opaque middleware: the MCP protocol lives inside Marta API

The Marta Agent connects through the Native MCP Server (POST /mcp, Streamable HTTP / JSON-RPC). Bearer authentication with connector token marta_conn_*, server name marta-agent, audit device type AI_CONNECTOR.

An optional stdio bridge is available for Claude Desktop. From the admin you download the skill pack and the filtered OpenAPI for ChatGPT Actions. Run an MCP smoke test without leaving Settings → AI Agent.

MCP inside Marta API, with a revocable token.

Connectors and channels

One token, two scopes, three ways to connect Claude and ChatGPT

Da Settings → AI Agent you create the connector. Choose the scope read (read-only) or read_write (read + mutations with dry-run). The plaintext token is shown only once: the database keeps the SHA-256 hash. You can revoke it at any time.

Guided setup for Claude Desktop (Mac and Windows), and guides for ChatGPT Plugin or Custom GPT. Other MCP clients that support Streamable HTTP or the stdio bridge.

Claude Desktop · Mac stdio bridge
Guided installer from Terminal. Gatekeeper blocks downloaded .command files.
Claude Desktop · Windows .cmd
.cmd installer + an MCP JSON snippet for developers.
ChatGPT Plugin Direct HTTP
Plus/Pro with Developer mode. Native MCP Streamable HTTP.
Custom GPT · Business / Edu OpenAPI
Actions on /ai/openapi-agent.json + skill zip.
Operational security

No write without a snapshot, a dry-run and explicit confirmation

The flow is mandatory: context first, then the numbers, then the draft. Mutations start with dryRun=true. Apply requires dryRun=false e confirm=true. About 30 mutations/hour per connector. Paths are blocked even with invoke_rest: AI connectors, OAuth, passwords, /mcp. Every write is audited with origin mcp o actions.

MCP tools

Reads, drafts and writes with the same permissions as the operator

Read tools cover the snapshot, KPIs, sales, customers without PII, promotion performance and the knowledge base. With scope read_write the agent updates settings, wallets and tiers, and creates promotions in status INACTIVE and only switches ACTIVE / INACTIVE.

invoke_rest is the bridge to the admin REST API: the same operator ACLs, no MARTAIA in v1, and no security paths.

get_program_snapshot
Detail, wallets, tiers, promotions, KPIs
read
query_sales / query_customers
Aggregated sales and customers, without PII
read
draft_promotion
Italian input → workflowConfig
write
create_promotion
Default status INACTIVE
write
upsert_wallet / upsert_tier
Creates or updates wallets and categories
write
invoke_rest
GET/POST/PUT on the admin REST API, operator ACL
bridge

Two AI layers

Marta Agent is not MARTAIA: one speaks from Claude, the other lives in Admin

MARTAIA is the intelligence inside Admin: Analyst, Promo Generator, Config AI. The Marta Agent works on the platform from external clients, with dedicated tokens, scope and dry-run. Same program, two ways of working — and an intentional split: the v1 agent does not call MARTAIA.

Want the AI inside the back office? Go to Marta AI — MARTAIA.

Marta Agent (MCP)MARTAIA
WhereClaude, ChatGPT, MCP clientsInside Admin
What it doesConfigures the program, promotions, wallets and RESTAnalyst SQL, Promo AI, Config AI
AuthToken marta_conn_*Operator session
WritesDry-run + mandatory confirmationDirect from the Admin context
Two distinct AI layers. The v1 agent does not call MARTAIA.

Native MCP, revocable tokens,
writes only after confirmation

MCP Server

Native POST /mcp endpoint. Streamable HTTP, JSON-RPC, server name marta-agent. No third-party proxy.

NATIVE

Scope and tokens

Read or read_write. The token is visible only once, the hash is stored in the DB, revocation is immediate, and the IAM permission is ai_connectors.

CONNECTORS

Dry-run

Every mutation is simulated first. Apply only with dryRun=false and confirm=true. About 30 applies per hour.

SECURITY

Claude and ChatGPT

Desktop on Mac/Windows, MCP Plugin and Custom GPT Business/Edu. Skill zip and OpenAPI you can download from the admin.

CHANNELS
All features

AI Agent and MCP: configure Marta in conversation

Native MCP Server (POST /mcp) with Bearer authentication and an AI_CONNECTOR audit.
Connectors with a read or read/write scope, a one-time token and immediate revocation.
Dry-run and confirmation required on every mutation, with a cap of about 30 applies per hour.
Dedicated tools: snapshot, reports, promotion draft, wallet, tier and promotion status changes.
REST bridge invoke_rest with the same permissions as the operator, and security paths blocked.
Guided setup for Claude Desktop (Mac/Windows) and guides for ChatGPT Plugin / Custom GPT.
Skill and OpenAPI for the agent, downloadable from the admin, plus an in-app MCP smoke test.
No PII on customer queries, and no call to MARTAIA from the v1 agent.

Connect Claude and ChatGPT to your program

Request a demo and we can look together at the token, the dry-run and a promotion draft via Marta Agent — on the same program you will use in production.