Get started

Connect Claude Code and other MCP clients

Let Claude Code, Cursor or any MCP client ask for approval before it acts, with one command and a sign-in, no API key.

About 3 minutes

Transcript (30 seconds)

Connect Claude Code to promptcompute with one command.

Add the MCP server, then sign in with /mcp. No API key needed.

When Claude wants to do something risky, it asks first.

You approve, edit or reject from the inbox, email or chat.

Claude runs only the approved arguments.

Before you start

A promptcompute workspace
Claude Code, Cursor or another client that supports remote MCP servers over HTTP

In your project folder, add our official MCP server.

claude mcp add --transport http promptcompute https://mcp.promptcompute.ai/mcp

Sign in
In Claude Code, type /mcp, choose promptcompute and sign in in the browser window that opens. Pick the workspace and environment the agent should use.
Ask for something risky
Ask Claude to do something that needs a person, such as a refund or a production change. It calls request_approval and waits for the decision.
Decide
Approve or reject from the inbox, email, Slack or Teams. Claude gets the decision with final_args, the arguments it is allowed to run.

Clients configured with JSON take the same URL. In Cursor, add it under Settings, MCP.

{
  "mcpServers": {
    "promptcompute": {
      "type": "http",
      "url": "https://mcp.promptcompute.ai/mcp"
    }
  }
}

The situation: Your support agent can issue refunds, but anything over $50 needs a person.
What to do: Connect it over MCP and publish a policy that auto-approves refunds up to $50.
What happens: Small refunds run at once; larger ones wait for an approver and run with the approved amount.

The situation: You want Claude Code to ask before git push or rm -rf.
What to do: Run npx promptcompute hook claude-code setup from the quickstart; it adds a PreToolUse hook with a gate file.
What happens: Matching commands pause until someone approves them in promptcompute.

Requests from the agent appear in your inbox, and the agent continues only after a decision.

It worked when

Example output: What the agent receives after approval

{
  "status": "approved",
  "final_args": { "order_id": "A-1001", "amount": 80 },
  "decided_by": ["Dana R."],
  "ledger_entry_id": "led_01j..."
}

The sign-in window does not open: Run /mcp again and choose Authenticate. Some terminals print the link instead; open it yourself.
The tool call times out: Clients without the MCP Tasks extension wait up to 50 seconds, then poll get_approval. The request stays open until it is decided or expires.
Calls fail with forbidden: Your organization may have an IP allowlist. Ask an owner to add your network under Org security.