Create Accountor Login
Docs

AI Agents

Part of the MountDev Ops Workspace documentation. Updated November 17, 2026

MountDev Ops Workspace documentation | Updated November 17, 2026

Connect Claude, or any AI app that supports MCP connectors, to your workspace and work in it by asking. The agent works as you, with your permissions, and never more.

Connecting an Agent

  1. Open My AI Agents in the Account menu (the Account button at the bottom of the menu), which replaced the AI Agents page (see MountDev Connections). A connector added earlier at an address with no company in it no longer connects: remove it in your AI app and add the address this page shows. The page says which company the address connects to and suggests a connector name, such as MountDev Ops: Acme Test Co. Each company has its own address, so to connect another company, use Switch Company first.
  2. In Claude, open Settings, then Connectors, choose Add Custom Connector, and paste the address. Other AI apps that support MCP connectors work the same way.
  3. Your browser opens the approval screen. It shows which app is asking, which company it will work in (in the title), what it can and can never do, and the site it sends you back to. Sign in first if you are asked to. If you are signed in to another of your companies, choose Continue With and the company's name to switch to the right one.
  4. Choose Approve or Decline. Nothing is connected until you approve. If you added the address in your AI app before choosing what it may do, the approval screen asks you to tick what it may do first, then shows the same review.

Each connection is tied to one company and to the person who approved it. An older address that has no company in it is retired and answers "This address is retired. Open My AI Agents in your company and use the address it shows." Use the address on My AI Agents instead.

Each person connects their own agents. An agent you connect works as you, not as anyone else.

What an Agent Can Do

  • See and change the records you can see and change, with exactly the same checks as your screens.
  • Read a login field's address and user name. The password comes only when the agent asks for it on purpose (core_records_login_reveal), only on records you can change, and each time is written in the Activity Log as done by an AI agent.
  • Read Home, your apps, modules, the people list, and usage.
  • Add modules and fields. Anything an agent adds is saved turned off until a person in your company turns it on.
  • In Helpdesk, read tickets and add internal notes.
  • In Forms, read your forms, their submissions, and Forms Settings (never the secret key). Building and changing forms stays with people.
  • If you are an owner or admin, read your automations and their runs, draft new ones, and turn them off. Anything an agent saves is turned off, and only a person turns an automation on.

What an Agent Can Never Do

  • Change people, roles, or anyone's access.
  • See or change billing, API connections, API keys, webhooks, or other AI agents.
  • Turn an automation on.
  • Export, import, or delete modules and fields, or export everything.
  • Reply to a customer by email, or send any email to a client. That always stays with a person.
  • Approve a proposal to be sent. Only a person approves; an AI assistant may pass on a named person's own approval, as that person and with every check theirs, and it shows as theirs "via" the assistant (Approve and Schedule).
  • See prices you cannot see yourself (Who Sees Prices).
  • Do anything you cannot do yourself.

Every change an agent makes goes through the same checks as the screens and is marked as made by an AI agent in the activity log and in each record's history. Changes made by AI agents are also counted on their own in Usage.

Seeing and Disconnecting Agents

My AI Agents lists the agents you connected (action core.agents.list): their name, the site they belong to, when you connected them, and when they last worked. Owners and admins see everyone's connected agents.

Choose Disconnect (action core.agents.revoke) to end a connection. It stops working on its very next request. Changes it already made stay, marked as made by an AI agent. You can connect it again at any time. Owners and admins can disconnect anyone's agent. When a person's access is turned off, their agents stop working too.

Your AI Agents' Keys

Owners and admins find Your AI Agents' Keys in Integrations, Apps and AI (the My AI Agents page keeps a link to it): the keys your company's AI agents use, the same for every company. Each is one of your company's connections (actions core.connections.save, core.connections.check, and core.connections.delete), stored encrypted and never shown again. AI agents never see or change them.

  • Add the Claude API Key: the key your company's AI agents use to think and write, from the Claude Console. It starts with sk-ant-; an Admin API key is refused. Its usage is billed to your Claude account. Check Now asks Claude which models the key may use.
  • Add an MCP Server: an outside MCP server your AI agents may use, such as a website's MCP connector: a name, its whole address, and a token if it needs one. Some servers keep their key in the address, so the whole address and the token are stored encrypted together, and only the address's start (such as https://mcp.example.com) is shown. Add as many as you need. The address must be https on the standard port, and never on a private network. Check Now asks the server to start an MCP session.

Your MCP Servers Through One Connection

Each person's AI agent connection to the workspace also carries the tools of your MCP servers that the person was given, so one connection reaches everything, however many servers you add. Owners and admins choose who may use each server, and which of its tools, in Settings, Permissions, MCP Servers (actions core.mcp.access.list and core.mcp.access.set):

  • No access, Every tool, or Some tools for each person on each server. A change reaches their AI app on its next request, with no reconnecting. Earlier access is kept on record, never deleted.
  • Each tool's name starts with mcp_, then the server, so two servers' tools with the same name are never confused.
  • The server's address and token stay encrypted in your workspace. People connect with their own sign-in only.
  • A tool that could send a message or an email, post, invite, or charge, pay, refund, or transfer money is held back: it shows in the list but can never be given to anyone. AI agents never email your customers or move money.
  • Every use of a tool is in the Activity Log, marked as made by an AI agent, with the tool and the server (never what was asked or answered). Check Now on a server lists its tools again.

If your company has Managed MCP from MountDev, MountDev sees counts only: how many people have access, how many servers there are and answer, how many tools are offered, and how many calls were made.

How the Connection Is Protected

  • The agent signs in with OAuth 2.1 and PKCE. It never sees your MountDev password or your workspace session.
  • Its access lasts an hour at a time and renews itself. Each renewal works once: if an old one is ever used again, the connection ends at once and must be approved again.
  • Every request is checked against your company's own record of the connection, so a disconnected agent is refused immediately.

Ask Us Anything

We’d love to hear from you!