Connect your AI to Gonzo (MCP)

What you can connect

Gonzo has one list of tools, offered three ways:

MCP, the Model Context Protocol, for Claude Desktop, Claude Code, Cursor and a ChatGPT remote connector.
A plain web API with an OpenAPI document, for a custom GPT's actions, Zapier, Make, n8n or curl.
These pages, so a person can read what the tools do before connecting anything.

Any AI can connect as a client, and doing so does not change which models Gonzo's own agents use. The tools themselves call no AI model. They return written facts, arithmetic and simulated results.

The tools, and what each returns

Tool	What it gives your AI
list_systems, get_system	The nine fixed-price systems: price, days to live, what each connects to, and what is built and what is not
price_list	The published price list in dollars, or the rupee list excluding GST, with its source
roi_estimate	The cost sum from the calculator page on your own numbers
list_guardrails	The rules every agent runs under: AI disclosure, consent, daily cap, kill switch, audit log
run_sandbox_demo, verify_demo_audit_chain	A deterministic, clearly labeled simulation, and a check that its audit log catches an edit
get_country_facts	Messaging, data-protection and cross-border billing facts for a market, only where a primary source is cited
answer_doubt	A written answer to a buyer's question, matched by words from a fixed list. A weak match returns no answer
how_to_request_free_audit	How the free audit works. It never takes a web address or an email
compare	A dated, sourced comparison with an alternative, including when the alternative wins
tenant tools (demo only)	Leads, conversations, bookings, receivables and ad numbers of a simulated business, and proposals that wait for a person

A reply that comes from the sandbox is marked simulated: true, and every reply names its source.

What no tool will do

They never write anything or take an irreversible action. A tool that could send, invite or reply only files a proposal, which waits until a named person on the client's side approves it somewhere else. No tool can approve.
They never touch Gonzo's own finance desk: no negotiation, no quote, no invoice, no price limits.
They never accept a password, a token or a key as an argument. A call that carries one is refused.
They never message a stranger. A WhatsApp reply is only possible inside a conversation the contact started.
They never call an AI model, and they never claim a client result. There is none.

Connect a client

A person at Gonzo issues the key, and there is no self-service sign-up. Replace https://your-gonzo-endpoint and gz_... with the values you are given, and treat the key like a password.

Claude Code

claude mcp add --transport http gonzo https://your-gonzo-endpoint/mcp --header "X-API-Key: gz_..."

Cursor (.cursor/mcp.json)

{ "mcpServers": { "gonzo": { "url": "https://your-gonzo-endpoint/mcp", "headers": { "X-API-Key": "gz_..." } } } }

ChatGPT, remote MCP connector. Add a custom connector with the URL https://your-gonzo-endpoint/mcp. It needs a public HTTPS address and a key sent as Authorization: Bearer gz_.... Check which sign-in methods the connector screen offers: this service takes an API key only, so a connector that insists on OAuth cannot use it.

Claude Desktop. The simplest form is a local process that Claude launches, set up from the developer docs and the repository notes. Claude Desktop's remote connectors may expect OAuth, and this service does not offer it.

A custom GPT. In the GPT builder, import the action from https://your-gonzo-endpoint/openapi.json?tiers=public_read. Set authentication to an API key with the header name X-API-Key. The tiers=public_read filter leaves out the demo tenant.

Zapier, Make, n8n. Use the product's HTTP step with the method GET, the URL https://your-gonzo-endpoint/v1/tools/roi_estimate?staff=2&hours_per_week=10&hourly_cost_usd=35, and the header X-API-Key: gz_....

curl

curl -s -H "X-API-Key: gz_..." "https://your-gonzo-endpoint/v1/tools/list_systems"

These steps are written from memory of each product's settings and have not been followed live. Check each product's own screen.

What has been tested, and what has not

Tested by our own tests, offline: the tool list and its rules, the key check that refuses every call when no key is set, the rate limit, the refusal of credential-shaped arguments, the OpenAPI document's structure, web API round trips against a local test server, the demo giving the same answer every time, and every call landing in a hash-chained audit log.
Tested against the MCP software development kit's own server code: building the server and listing and calling tools. A full client session has not been run in an environment where it works.
Not tested: ChatGPT, Claude Desktop, Claude Code, Cursor, Zapier, Make, n8n, a real network connection, a public deployment.
Not built: OAuth, real tenants, a client approval screen. The demo tenant is a simulation with made-up people.

See the developer docs for the web API, and Trust and method for the rules every Gonzo agent follows.

Connect any AI to Gonzo's tools.