Get started with PromptOps and MCP

PromptOps is the AI assistant in the console, in beta. Tell it what you need, in English or Arabic, and it does the work in the workspace you are in: it looks things up, explains what is failing, and creates or changes what you ask for, asking first. It acts as you, with your role, so it can never do more than you can. The same tools are open to your own AI agent, such as Claude Code or Cursor, over the MCP endpoint.

Before you start

• To turn PromptOps on and set what it may do: the workspace's Owner or Admin role. Everyone in the workspace can then use it, within their own role.
• For MCP: an MCP client, and an API token or a browser to sign in with.

1. Turn PromptOps on

PromptOps is off in every workspace until an owner or admin turns it on. Open the workspace's Settings → PromptOps and switch on PromptOps for this workspace. Switching it off there stops it at once, for everyone in the workspace.

2. Set what it may do

On the same page, What PromptOps may do is the permission matrix. Its rows are areas, such as instances, Kubernetes or load balancers, and its columns are View, Create, Change and Delete. Each cell is one of:

Every workspace starts with looking set to Run and every change set to Ask first. Reading database rows, app logs and container logs asks first too, because what it reads is sent to its AI provider. Deleting, firewall changes, database writes and changes inside a cluster always ask first, in every workspace. Reset to defaults goes back to the starting matrix. See The permission matrix.

3. Ask it

Click PromptOps in the top bar, or press ⌘J on a Mac or Ctrl+J elsewhere. Say what you want, with the names you see in the console:

• "What's running in this workspace, and is anything failing?"
• "Create a private PostgreSQL instance called orders."
• "Scale the default pool of the shop cluster to 3 workers."
• "Why did the last deploy of storefront fail?"

When it needs your approval, a card shows the action and its values: Approve runs exactly that, Decline cancels it. The mode beside the message box decides how it asks: Ask, the default, stops at every Ask first; Auto runs low-risk actions that match your request; Plan only looks, and ends with the changes it would make. Think harder plans a complex request more carefully. On the page of something unhealthy, Diagnose with PromptOps writes the request for you.

PromptOps never touches your account, billing, the workspace's members, API tokens or secrets such as passwords and kubeconfigs: those stay in the console.

4. Connect your own agent over MCP

The endpoint is https://api.sahabti.com/mcp. The workspace's API tokens page has an MCP panel with a snippet for each client. Two ways to sign in:

• OAuth: add the endpoint with no token. Your client opens a browser to sign in, and you choose the one workspace it may act in. It acts as you, with your role there.
• A token: create one on API tokens with New token, with only the actions the agent needs, and send it as the x-api-key header. Best for an agent that runs unattended.

Claude Code, with a token:

claude mcp add --transport http sahabti https://api.sahabti.com/mcp --header "x-api-key: <token>"

Ask the agent to call whoami first: it names the one workspace the connection acts in and what it may do there. A connection reaches one workspace only, and every call runs the same permission checks as the console.

What it costs

Each person has a daily number of PromptOps requests, and a message with Think harder counts as several. Owners and admins can buy more for the workspace under Settings → PromptOps → Buy more. See Daily requests and buying more, Limits and the pricing page. MCP calls run the same routes as the console, with nothing extra to pay for the connection.

First checks

• PromptOps appears in the top bar, and "How much of our quota is left?" gets an answer.
• A change asks you first, as the matrix says.
• Over MCP, whoami names the workspace you meant.

Next steps

• PromptOps: modes, confirmations, privacy, limits and troubleshooting
• MCP for AI agents: every tool, OAuth and tokens, and what an agent can't do
• API tokens: least-privilege tokens for scripts and agents