Skip to main content
Karzoun MCP is a thin bridge: one MCP tool = one public GraphQL operation. No hidden APIs, no extra abstraction layer.

Request path

  1. The agent picks a tool (e.g. customerDetail) and passes arguments.
  2. The MCP server builds a GraphQL query or mutation from the committed tool manifest.
  3. The server POSTs to your tenant GraphQL URL with x-app-token.
  4. The result is returned as JSON text to the agent.
The same app token permissions apply — if the app cannot run an operation in the Playground, MCP cannot run it either.

Tool naming

Tool names match GraphQL operation names exactly: Browse the full list in the tool catalog. Field-level documentation lives in the GraphQL reference.

Limiting tools

Large manifests can overwhelm smaller models. Set a prefix to expose only what you need:
Only tools whose names start with customers are registered (e.g. customers, customersAdd, customerDetail).

Response size

Responses are capped at 512 KB by default (KARZOUN_MCP_MAX_RESPONSE_BYTES). If a list query returns too much data, the tool fails with a clear error — narrow page / perPage or filter arguments, same as pagination in GraphQL.

Errors

Failed tools return JSON with an error field:
HTTP failures, GraphQL validation errors, and oversize responses all surface this way. Teach your agent to read the message and adjust arguments.

Stdio vs hosted

Hosted mode also supports optional x-subdomain when routing through multi-tenant gateways. See Hosted MCP.

What MCP does not do

  • Stream inbox real-time events — use webhooks
  • Register MiniApps — see MiniApps
  • Bypass permission checks — apps are scoped in Developer → Apps
Last modified on August 8, 2026