Setup & Connection

Connect your AI tool to Prediko and ask about your inventory, forecasts, and purchase orders in plain English — “which SKUs run out this month?”, “draft a PO for my best sellers”. The Prediko MCP server works with ChatGPT, Claude, and any other client that supports the Model Context Protocol, and exposes a curated tool surface covering inventory planning, purchasing, demand forecasting, and operational insights.

Connection URL

https://mcp.prediko.io/mcp

Note

https://mcp.prediko.io/mcp is the protocol endpoint and only speaks MCP. If you open it in a browser you’ll be redirected here to the docs.

MCP access is available on every Prediko account — there is nothing to request or switch on. Sign in with your Prediko login when your client asks.

Connect your AI tool

Claude (web, desktop & mobile apps)

1. Go to Settings → Customize → Connectors and choose Add custom connector.
2. Paste https://mcp.prediko.io/mcp and confirm.
3. Click Connect and sign in with your Prediko account.

Available on all Claude plans. On Team/Enterprise an org Owner adds the connector first (Organization Settings → Connectors), then each member connects individually.

ChatGPT (web)

1. Enable developer mode: in Settings, open Advanced settings (under the Apps or Connectors tab, depending on your version) and toggle Developer mode. (On Business/Enterprise an admin must first allow custom MCP connectors under Workspace Settings → Permissions & Roles → Connected Data.)
2. Go to Settings → Connectors → Create, name it “Prediko”, paste https://mcp.prediko.io/mcp, and pick OAuth for authentication.
3. Sign in with your Prediko account when prompted.

Custom connectors require a paid ChatGPT plan (Plus, Pro, Business, Enterprise, or Edu).

Any other MCP client

Prediko works with any MCP-compatible tool. Point it at https://mcp.prediko.io/mcp (Streamable HTTP transport) and complete the OAuth sign-in on first connect — no API keys needed.

For developers

Claude Code

claude mcp add --transport http prediko https://mcp.prediko.io/mcp

Cursor / VS Code (mcp.json)

{
  "mcpServers": {
    "prediko": {
      "url": "https://mcp.prediko.io/mcp"
    }
  }
}

Codex (~/.codex/config.toml)

[mcp_servers.prediko]
url = "https://mcp.prediko.io/mcp"

Authentication

The server authenticates with OAuth 2.1, backed by Auth0. Every tool call resolves the calling user’s tenant from the validated JWT — there is no API-key auth on the MCP surface (that’s the REST API).

1. Your MCP client runs the OAuth 2.1 authorization-code flow against Prediko’s Auth0 tenant. Sign in with your Prediko account and approve the requested scopes on the consent screen.
2. The client receives an access token (JWT) and attaches it as a bearer token on every MCP request. Tokens are verified server-side against Auth0’s JWKS (RS256).

Scopes

Scopes are advertised on the connector’s protected-resource metadata so the consent screen can show what the agent is asking for:

Scope	Grants
read:inventory	All read / search tools and insights
write:purchase_orders	Create / update / receive POs, and PO export
write:sku_settings	update_sku_settings

Note

Scopes are for consent-screen transparency. The real authorization boundary is the tenant resolved from your JWT plus backend permissions — you can only ever read or write data belonging to your own tenant.

Actor attribution

For write tools, the gateway calls Prediko’s backend with a shared service identity but forwards your identity alongside it. Backend audit records therefore attribute each change to you, the real caller — not to a service account. See Write Semantics for the audit_id returned on every write.

Tool surface

The server advertises 9 read tools and 4 write tools. See the full Tool Reference for what each does and its side effects.

Reads	Writes
find_entity, list_entities, get_entity	create_purchase_order
search_inventory, search_purchase_orders, search_forecasts	update_purchase_order
get_insights, get_change_history, export_purchase_order	receive_units, update_sku_settings

Every write tool accepts dry_run=true to preview its effect without persisting — see Write Semantics.