Run Gemini inside your workflows on your own key

Google Gemini in MadMax is a model provider for the AI agent steps inside your workflows. Paste an AI Studio key, bind it to an agent step, and that step runs on a Gemini model billed to your Google account. The rest of the workflow, from the trigger to the approval to the apps it writes to, does not change, and the agent keeps its full task list (custom prompt with a JSON output schema, extract, classify, summarize) and its tools: calculator, browser, sandboxed code, your sub-workflows, any operation of a connected app, plus conversation memory and knowledge-base retrieval. You still describe the automation in plain English and MadMax builds and verifies it, judging each agent's output on real upstream data rather than the status code. Gemini only changes which model answers when a bound step runs; without a provider bound, the step runs the platform default and the model field is ignored, so you can build first and switch models where it matters.

What teams build

• Employees email receipts to a workflow's inbound address, a Gemini-bound extract step reads the attachment and returns merchant, date, total and category, a store upsert keeps one row per receipt, and a Friday schedule emails each manager the week's list.
• A webhook delivers each help-desk ticket, a Gemini classify step sorts it into billing, technical, account or other, a second agent writes a two-line English summary of a ticket written in any language, and a switch posts it to the right Slack channel.
• A hosted form takes a topic and a list of source URLs, a Gemini custom-prompt agent with the browser tool reads them and writes a one-page brief, and an approval emails the editor a one-click approve or reject before the brief is appended to a Google Doc.

How it connects

Paste an API key from aistudio.google.com/apikey (it starts with AIza; it is a personal AI Studio key and no Google Cloud project is needed) into Settings → connectors → Google Gemini, or add it from an agent step's Provider picker on the canvas. The key routes the agent steps you bind it to through Google on your own account, and MadMax does not meter those calls. The canvas chat that builds and edits workflows stays on the platform model. The live web-search tool is a platform-model feature a Gemini-bound step ignores; browser, code, workflow and connected-app tools keep working. Rotating the key is a reconnect in place. See managing connections.

Related

• AI settings and models
• AI agent step
• Extract task
• Managing connections
• Claude on your own key