# Connect ChatGPT to Your Product with Peeve (MCP)

> Let users connect ChatGPT and OpenAI agents to your product over MCP with Peeve: capabilities become typed tools, scoped per user. On every plan.

Canonical: https://peeve.ai/integrations/chatgpt/

If your product runs Peeve, your users can connect ChatGPT and other OpenAI agents to it over MCP. They point the agent at your scoped Peeve endpoint, and it receives your product's capabilities as typed tools with a stated effect, scoped to exactly that user's permissions, with writes confirmed. The agent channel is on every Peeve plan.

## How your users connect ChatGPT

1. **Turn on the agent channel** - The agent channel is on every Peeve plan. ChatGPT and other OpenAI agents connect over MCP, the same protocol and registry Peeve serves to any agent, so there is nothing new to build.
2. **Your user gets their endpoint** - Each user has a personal, scoped endpoint, mcp.peeve.ai/<workspace>/u/<contact>, minted per user and revocable by them.
3. **They connect it in ChatGPT** - The user adds that endpoint as an MCP connector in ChatGPT's connector settings, the same way any custom MCP server is added.
4. **ChatGPT calls your tools with stated effects** - The agent discovers your capabilities as typed tools. Reads run live; writes, like issue_refund, state their effect and run only after approval, inside an allow-list, written to a signed audit log.

## Questions

### How do my users connect ChatGPT to my product?

Your users add their personal Peeve endpoint (mcp.peeve.ai/<workspace>/u/<contact>) as an MCP connector in ChatGPT's connector settings. ChatGPT then discovers your product's capabilities as typed tools and calls them on the user's behalf, with writes confirmed first. The agent channel is on every Peeve plan, so nothing needs upgrading to enable it.

### What can ChatGPT do once connected?

ChatGPT gets your capabilities as typed tools with a stated effect, for example list_orders, download_invoice or change_plan, the same registry a human sees. It reads live data to answer and takes write actions behind a confirmation step, with exactly the connecting user's permissions, rate-limited and written to a signed audit log.

### Is connecting ChatGPT safe?

Yes. ChatGPT acts with exactly the calling user's permissions, each grant is scoped and revocable, and any state change is confirmed and audited. The model never holds a credential: it requests a capability by name, so an injected instruction can request a capability but can never compose a raw request or reach a key.

### Do ChatGPT tool calls over MCP cost extra?

No. Agent tool calls over MCP are logged but not billed today, and the agent channel is on every Peeve plan, with grants limited by plan (5 per user on Starter, 20 on Growth, unlimited on Scale). If the billing model changes you get a full billing cycle's notice, never a retroactive charge.
