Guide

Connect ChatGPT through MCP

Preview guide

Paste into an AI assistant to help with integration. Documentation and examples only; no real credentials.

Preview guide

Connect ChatGPT through MCP

Direct answerConnect and authorize, then query documentation. Confirm each paid call or key change.

Updated · Reviewed

Connect this site's MCP server in ChatGPT

Once the site administrator enables MCP, add its address as a custom MCP connection in ChatGPT. Availability depends on your ChatGPT account and workspace settings. Ask the administrator for the pre-registered OAuth client configuration.

Connect and authorize

MCP address: https://api.tu-zi.com/mcp.

  1. Add this address using the custom MCP connection interface and the administrator-provided client configuration.
  2. Sign in on this site, review the account and requested scopes, and approve authorization.
  3. Return to ChatGPT and ask it to search image API documentation, inspect your usage, or generate content using one of your API Keys.

Never paste your password, account system token, or full inference API Key into chat. Select an owned API Key ID; the server retrieves the credential and uses the existing gateway.

Initial capabilities and limits

  • Documentation: overview, search, full parameters and examples, source URLs and revisions.
  • Account: own balance, model catalog and displayed prices, paginated usage logs for a range of up to 31 days.
  • API Keys: list settings, create, partially update and delete. View full keys only on the website.
  • Text: non-streaming POST /v1/chat/completions.
  • Images: POST /v1/images/generations, requesting URL output; retrieve asynchronous results by request_id.

The initial release exposes 15 tools by default and excludes video submission and polling. Video API documentation remains searchable. Video tools appear only after the administrator separately enables and validates them. Use the connected server's current tool list.

The model must support the chosen protocol. Account and key permissions, limits, balances and channel availability still apply. Documentation for other protocols is searchable, but not every protocol is an execution tool. Audio, realtime WebSocket, multipart image editing and administrative operations are outside this initial release.

Confirm and retrieve results

Each key change or paid call first creates an action awaiting confirmation. Open the site's approval link, inspect the account, model, API Key ID and request parameters, then confirm within 15 minutes. Return to the assistant to execute the action.

Model execution initially returns an executing state; use get_action to retrieve the saved result. Repeated calls for the same action do not execute it again. For executing or unknown outcomes, inspect usage logs and task results before taking further action. Do not automatically create an identical paid action. Task acceptance does not mean generation has completed; continue polling with the task ID.

If an output exceeds MCP response limits, inspect it on the website instead of generating again. Account data and operation results are limited to the authorized connection.

Disconnect and troubleshoot

Revoke authorization on the MCP connections page. Requests already running may finish and incur charges. Disconnecting does not delete existing API Keys.

For connection issues, check that MCP is enabled, its public HTTPS endpoint is reachable, the callback URI is registered exactly, and the account and required scopes are valid. Documentation tools respect the site's document access policy, including login requirements.

Security limits and retention

Paid calls require an administrator-approved model and an owned, finite-quota API Key within the site's MCP limit. MCP cannot enable unlimited quota. Text calls accept text-only messages, with an output limit of 8192 tokens. Images are capped at four per call. If the administrator separately enables video tools, videos are capped at twelve seconds and reference-media URLs require an explicitly trusted HTTPS host. Model-specific limits may be stricter.

Only one MCP operation may execute per account at a time. The default limit is 30 executions per UTC day, including key changes, and is configurable by the administrator. Request-size and key-quota limits are not a price quote; the gateway settles actual charges. The account tool exposes the current MCP policy.

Account data and results are sent to the authorized assistant. MCP stores operation content encrypted, stops serving it after seven days, and removes it in background batches. Existing gateway logs, providers and chat platforms follow their own retention rules. Changes to key routing or permissions require a newly reviewed operation.

Use cases

  • Find API integration documentation
  • Approve model calls for your account

API protocols

  • /mcp

FAQ

Should I give the assistant my full API Key?

No. The assistant uses a key ID; full credentials remain on the server.

Why confirm on the website?

Key changes and paid calls affect your account. The page shows the specific action; approval alone does not execute it.

Should I regenerate after a timeout?

Do not create another identical paid action. Check the original action, usage and task results first to avoid duplicate charges.

Official sources

  1. OpenAI MCP authentication Official