Docs
Developer

MCP tools

Public MCP tool surface, OAuth, and governed execution for agent clients.

Overview

Numezis exposes a remote MCP server for Claude, ChatGPT, and other MCP clients. tools/list publishes callable domain façades. A façade runs one governed API operation: reads stay reads, and mutations still follow module activation, permissions, and approval policy.

End-user setup (connectors, OAuth screenshots) lives in the MCP product guide. This page documents the integration contract for builders.

Endpoint and transport

ItemValue
URLhttps://www.numezis.com/mcp
TransportStreamable HTTP (JSON-RPC 2.0)
AuthOAuth 2.0 with Dynamic Client Registration

Send portal context through the authenticated session established at OAuth time. Tool execution uses the same AccessMode rules as the company portal (/c).

Tool discovery

After authorization, call tools/list. The list is paginated (nextCursor). Each entry is one callable façade, named numezis.<domain>.<verb> or numezis.<domain>.<family>.<verb>. The list does not include numezis.load_domains, a retired alias, or a generic dispatcher such as sales.query. A second call returns the same names.

Each tool includes:

  • name — the façade to pass to tools/call
  • description — the consent sentence shown before the host asks the user
  • inputSchema — JSON Schema; a family façade takes resource plus args
  • annotations.title — localized title (Domain · Family · Verb)
  • annotations.readOnlyHint and annotations.destructiveHint — whether the call only reads, or can change records

Only façades for enabled modules and effective permissions appear for the connected user. _meta.toolAdmission is domain_facade_catalog.

Execution and approvals

  1. Client sends tools/call with tool name and validated arguments.
  2. Server runs policy checks (module, permission, company status).
  3. Critical tools may return an approval token instead of applying the mutation.
  4. User confirms in Numezis UI; client retries with the approval reference.

Do not bypass approval steps in automation—use non-critical read tools to preflight data, then surface approval UX to the user.

Errors

MCP errors mirror API semantics: authorization failures, validation errors, and rate limits are returned as structured RPC errors. Log correlation ids from responses when opening support tickets.