MCP server

Requirements

• A Pro plan. Free accounts can't create MCP keys, and a key on an account that leaves Pro answers every tool call with pro_required until the account is Pro again; the key itself keeps working once it is. See pricing.
• An MCP client that supports the streamable HTTP transport with a custom header: Claude Code, Cursor and VS Code all do.

Create a tenant key

Sign in, open Dashboard → Keys, and click Create MCP key in the MCP access card. The full key is shown once; we only store a hash. Each account can hold up to 3 active tenant keys, so issue one per machine or per teammate and revoke any one without touching the others.

Tenant keys are separate from API keys. An X-API-Key can't authenticate MCP, and a tenant key can't call /v1/predict. Tenant keys can only be created, listed and revoked in the dashboard, never over MCP, so a leaked key can't mint more keys.

Connect your client

Send the key as Authorization: Bearer sf_tenant_…. The card on the Keys page fills these snippets in with your key.

Claude Code

claude mcp add --transport http siftfy https://api.siftfy.io/mcp \
  --header "Authorization: Bearer sf_tenant_…"

Cursor — ~/.cursor/mcp.json

{
  "mcpServers": {
    "siftfy": {
      "url": "https://api.siftfy.io/mcp",
      "headers": {
        "Authorization": "Bearer sf_tenant_…"
      }
    }
  }
}

VS Code — .vscode/mcp.json

{
  "servers": {
    "siftfy": {
      "type": "http",
      "url": "https://api.siftfy.io/mcp",
      "headers": {
        "Authorization": "Bearer sf_tenant_…"
      }
    }
  }
}

The Cursor and VS Code files hold the key in plain text. Keep them out of version control, or add the key to your user-level config instead of the project's.

Tools

17 tools. Every tool runs the same validation as the dashboard, so an agent can't put a site into a state the dashboard would refuse.

Read

Write

update_content_gate_site needs confirm only when it moves a site to enforce mode. New sites always start in monitor mode.

The confirm guard

A tool marked confirm: required changes nothing unless the call includes "confirm": true. Without it, the tool returns a preview of the effect so the agent (and you) can see what would happen first:

{
  "applied": false,
  "confirm_required": true,
  "effect": "Rotating the secret breaks siteverify on every server using the old secret until it is updated."
}

Things to ask

Plain requests work. The agent maps them onto the tools above.

• “Add Content Gate to the signup and contact forms on staging.example.com.”
• “How many predictions did we run each day this week?”
• “Create an API key called checkout-service.”
• “Move the production site to enforce mode once the review rate is under 2%.”
• “Deny these three solver IDs on the signup action for 7 days.”
• “Rotate the Content Gate secret for production.”

Errors

A missing, malformed or revoked key gets HTTP 401 with invalid_tenant_key before any tool runs. Revoking a key in the dashboard takes effect on the next request. A tool that fails returns an MCP tool error whose structured content carries a stable error code:

{
  "error": "pro_required",
  "message": "MCP access is a Pro feature. Upgrade at https://siftfy.io/dashboard/billing and this key works again unchanged.",
  "billing_url": "https://siftfy.io/dashboard/billing"
}

Limits and audit

MCP calls share your plan's per-minute budget (see rate limits), counted separately from /v1/predict traffic. One request carries one call; JSON-RPC batches are refused. Every write call is recorded with its key, tool and result. The MCP access card shows the latest 50; we store a hash of the arguments, never the arguments themselves.

Where to next

• Content Gate → what the site, policy and deny-list tools configure.
• Authentication → API keys for /v1/predict.
• Pricing → what Pro includes.