AI Integration is Marzi's MCP server. It does two things: your assistant reads any public Allegro data the moment it needs it — a listing, a category, a seller, a product card with every offer, one offer in full — and it keeps watching that data after the conversation ends, with monitors Marzi runs on its own schedule. Your products, the weekly index and the store analysis come on top.
Niche research, price checks, competitor watching, position reports — the assistant does them on today’s Allegro instead of guesses, and keeps the watch running. Pay per request; the first 5,000 are free; a cached repeat or a failed read costs nothing.
Your assistant reads Allegro the moment you ask: a search with sort and filters, a category, a seller's storefront and reputation, a product card with every seller on it, one offer at listing depth or in full, a card by barcode. Every card comes in one normalised shape with the page and time it was read.
Tell the assistant what to watch — a price, a stock level, a competitor's offer on a card, your own position for a search — with a rule and an interval from one hour to a day. Marzi runs the checks itself, keeps the history and writes an event only when the rule fires.
Copy the endpoint into Claude, ChatGPT or the Codex CLI and sign in when the browser opens; SDKs carry a personal token shown once. The wizard lists every function with a plain name — what is switched off is refused at the server — and the dashboard shows what was called, when and what it cost.
Every figure on the screen has the page it was taken from and the method printed next to it. These are the parts.
Search results with sort, filters and a cursor; category and seller listings; a seller's reputation; product cards with every offer, price with delivery, stock and buy box; a card by barcode; one offer in full — stock, GTIN, rating distribution, reviews by feature, description, warranty, returns. Read as a shopper sees them now, in one normalised shape.
Watch a price, a stock level, a competitor's price or stock on a card, or your own search position — with a rule and an interval from an hour to a day. Marzi runs the checks on its own schedule, keeps the history and writes an event when the rule fires; the assistant reads events back and you see them under the bell.
Your own list with the positions Keyword Tracker measures, and the weekly market index for sizing a niche — answered from Marzi's database, without opening Allegro.
The tracked products, their cards and positions turned into prioritised findings with the evidence quoted by field — a deterministic read, not a model's opinion.
Chat clients sign in through the browser; scripts carry a personal token that is shown once and stored as a hash. Anything not switched on in the wizard is refused by the server.
Every request is one line with its price at the moment of the call. A cached repeat and a failed read cost nothing; trial requests come first, then the prepaid balance, then the daily cap you set.
In colour: connected and verified end to end, with the setup steps in the cabinet. In grey: on the way — any client that speaks MCP over Streamable HTTP with a bearer token works today; these get their own guide once we have tested them ourselves.
ClaudeDesktop and claude.ai · custom connector, OAuth sign-inClaude CodeCLI · claude mcp add --transport http
ChatGPTCustom app in developer mode · Pro, Business, Enterprise/Edu (not Plus) · web only
CodexCLI · codex mcp add, OAuth or personal tokenOpenAI Responses APIMCP tool in the request, with a personal token
Any MCP clientStreamable HTTP and a bearer token — the official SDKs in Python and TypeScript
Perplexitycoming soon
Google Geminicoming soon
Ollamacoming soonAllegro data on demand and the same data tracked over time are the base; your products, the index and the store analysis come on top. Each function is switched on or off in the cabinet; what is off is refused at the server. Every answer carries the source page and the time it was read.
Live reads of Allegro’s public pages — what a shopper sees right now. Listings are cached for ten minutes, offers for an hour.
The same facts, watched on Marzi’s side after the conversation ends — from every hour to once a day.
Answers from Marzi’s own database — fast, without opening Allegro, as fresh as the last measurement.
Your tracked products, their cards and positions turned into prioritised findings with the evidence quoted.
The pair every ChatGPT connector needs.
No function names, no parameters. The assistant picks the call; Marzi opens the page, extracts the figures and hands them back as data — and checks that the digits in the answer are the digits it returned.
every function reads public pages or Marzi’s data; nothing is written to Allegro
every link is classified before it is opened; anything else is refused
sign in through the browser; no client id, no secret, no key to paste
shown once, stored as a hash, rotated from the cabinet in one click
a function switched off in the cabinet is refused at the server, whatever the client asks
priced at the moment of the call; the counting rule is shown in the cabinet
The same server behind Claude, ChatGPT, Codex, the OpenAI Responses API and any MCP SDK. Chat clients sign in through the browser; scripts carry a personal token from the cabinet.
claude mcp add --transport http marzi https://mcp.marzi.ai/mcp
# sign in when the browser opens, then /mcp to see the tools# Settings → Connectors → Add custom connector Name: Marzi for AI URL: https://mcp.marzi.ai/mcp Auth: OAuth — no client id, no secret
codex mcp add marzi --url https://mcp.marzi.ai/mcp codex mcp login marzi # or ~/.codex/config.toml with a personal token [mcp_servers.marzi] url = "https://mcp.marzi.ai/mcp" bearer_token_env_var = "MARZI_MCP_TOKEN"
{
"type": "mcp",
"server_label": "marzi",
"server_url": "https://mcp.marzi.ai/mcp",
"authorization": "<your personal token>",
"require_approval": "never"
} Endpoint: https://mcp.marzi.ai/mcp · Streamable HTTP · Read-only
Marzi says which is which on the screen itself. Here is the short version for AI Integration.
Ask what sells in a category, who sells a product and at what price with delivery, what a card's stock and buy box look like — and get the page's numbers, not a summary of the internet.
Put the data you care about under watch — competitors' prices and stock, your own positions — and let the assistant tell you what moved while you were shipping.
Wire Allegro into your own agents through the Responses API or an MCP SDK, with one token per account and a meter you can read.
Any public Allegro page, read the moment it is asked: a search with sort and filters, a category, a seller's storefront and reputation, a product card with every seller, price with delivery, stock and buy box, one offer at listing depth or in full, a card by barcode. Facts as Allegro shows them, with the time they were read; unknown is null, never zero.
Any of those facts: an offer's price or stock, a competitor's price or stock on a card, your own organic position for a search. A rule and an interval from an hour to a day; Marzi runs the checks itself, keeps the history and records an event when the rule fires.
Claude (custom connector) and Claude Code, ChatGPT (a custom app in developer mode — on Pro with read/fetch, which is all Marzi needs, and on Business and Enterprise/Edu; Plus and Free cannot connect one), the Codex CLI, the OpenAI Responses API and any MCP SDK with Streamable HTTP and a bearer token. The wizard shows the steps and the config for each. Clients we have not connected ourselves are not on the list.
Chat clients use OAuth 2.1: paste the endpoint, a Marzi page opens, you sign in, done — no client ID or secret. Scripts and SDKs use a personal token generated in the cabinet; it is shown once and can be rotated.
The account's rate per 1,000 requests, applied at the moment of the call. New accounts start with 5,000 free requests. A cached repeat — the same listing within 10 minutes, the same offer within an hour — and a failed read cost nothing.
A watcher the assistant creates in words — a link, a rule, an interval from one hour to a day. Marzi checks it on schedule and records an event when the condition is met; the assistant reads events back, and you see them under the bell in the cabinet. Each scheduled check is one request.
Yes, twice. The wizard switches individual functions on or off, and the server refuses what is off. The daily limit caps spend per day; when it is reached, calls are refused until tomorrow.
No. Every function is read-only: it reads public pages or Marzi's own data. Monitors are stored in Marzi, not on Allegro.
7 days, no card required. The whole index, your trackers and the MCP endpoint are live from your first sign-in.