Platform Agent

The Platform Agent is the AI assistant you chat with in the Major web app. Describe what you want and it plans the work, asks for approval where needed, and builds it: apps, agents, skills, and workflows. It can also query your connectors directly to answer questions about your data.

Example requests

• “How many deals did we close last quarter, by region?”
• “Build an app that tracks open support tickets from our Postgres database.”
• “Create an agent that posts a summary of new Salesforce leads to #sales every morning.”
• “Write a skill that explains how our billing tables fit together.”
• “Make a workflow that runs the lead-scoring agent whenever a new Attio record is created.”
• “Add a date filter to the revenue dashboard and deploy it.”

What it can do

How it works

Plans

For complex, ambiguous, or long-running requests, the Platform Agent tells you its approach before it starts, or asks a question when there are meaningfully different ways to do the work.

Questions

When the Platform Agent needs a decision from you, it shows a question card with one or more questions and a set of options for each. Some options include a visual preview, such as a layout mockup, so you can compare them. Pick your options and click Submit answers. You can also:

• Attach files to an answer - click Attach files under a question, or drop files on the card, to send them with that answer.
• Answer in your own words - type in the chat composer while the card is open and click Respond. Your message is sent as the answer, along with any options you already picked. Questions you didn’t pick an option for show as Answered in chat.
• Skip the questions - open the menu next to Respond and choose Skip questions and send to send your message as a new instruction instead.

Once answered, the card shows a summary of each question’s answer and attached files, and marks any you skipped.

Sandboxes

Apps, agents, skills, and workflows are each edited on a sandbox attached to the chat. Several can be attached at once, so one conversation can build an app and the agent it calls. A sandbox is held by one person at a time; if someone else is editing the same item, the Platform Agent tells you it is locked. Chats you start from an item’s editor are pinned to that item, and its sandbox starts automatically.

Tool cards

Each action the Platform Agent takes appears in the conversation as a tool card, so you can see what it read, queried, edited, or ran. Approvals, questions, and setup requests appear as cards you act on directly.

Approvals

When the Platform Agent wants to use a tool that needs your sign-off, and the chat is in Manual approvals or auto review hands the call to you, it shows a Permission request card with the tool name, the reason, and the arguments it will send. Choose: By default, read-only connector tools and app GET requests run without asking, and tools that write ask first. Actions such as running an agent, running a database migration, or deleting an app, agent, or workflow always ask. Manage your saved rules, or set always allow, ask, or always deny for any tool ahead of time, on the Tool Permissions page. When you chat with a specific agent, that agent’s tool permissions apply instead of yours.

Auto and Manual

The Approvals menu next to the send button sets who decides tool calls that would otherwise ask you: Every chat starts in Auto. In Auto, the reviewer looks at the call and the recent conversation and judges whether you would approve it. Calls that move your request forward run. A call is denied if it could leak data in a way you wouldn’t expect, has nothing to do with what you’re working on, or touches very sensitive data (such as writing to a production database) that you didn’t clearly ask for. When a call is denied, the Platform Agent is told to take a safer route or to stop and ask you. Auto mode still sends some things to you:

• Questions, connector setup and sign-in requests, and environment variable requests always go to you.
• If the review fails or the call is too large to review, you get a Permission request card.
• After three auto-denied calls in a row in a chat, the next call comes to you.

Auto only covers calls that would otherwise ask. Tools you set to always allow still run, and tools set to always deny stay blocked. Only the person who owns a chat can switch it between Auto and Manual. On a shared edit chat that no one owns, anyone with access can switch it.

Shell command screening

Shell commands the Platform Agent or an agent runs on Major-hosted machines, in its own workspace or in an app sandbox, run without asking. Before each one runs, Major screens it for attempts to attack Major’s infrastructure, such as reaching cloud-metadata addresses or internal services, sending the machine’s credentials elsewhere, or escaping the container. Ordinary development is never blocked, even when it’s destructive to the workspace itself. When a command is blocked, the agent is told so and continues without it. Connector calls get a similar check; see Call screening.

Models

Pick a tier - Smartest, Balanced, or Fastest - from the model picker when you start a chat. Each tier maps to the model your organization configured for it. The model is fixed once the chat starts, and a chat with an agent uses that agent’s model. Admins choose the model behind each tier in Settings > Models.

Conversation files

Files the Platform Agent writes in its working directory, such as notes, scripts, and intermediate results, persist for the conversation, so you can pick up where you left off.

Platform Agent and MCP

The Platform Agent uses the same tools and skills as the Major MCP server. If you prefer to work in Claude Code, Codex, ChatGPT, Cursor, or another MCP client, connect it to Major and you get the same capabilities there. The main differences:

Next steps