Build on the Lokrix AI GEO engine

Base URL

https://api.lokrix.de with a /v1 version prefix on every route.

Pagination

Offset pagination via limit + offset; responses return { items, total, limit, offset }.

Compression

Send accept-encoding: gzip, br — payloads are compressed over the wire.

Provenance & CI

Every scored payload carries a provenance badge (live / cached / predicted) and a 95% confidence interval.

Approval gate

Gated workspaces receive a 202 pending_approval envelope until an admin approves the action.

Rate & budget

Over-limit requests return 429; over-budget live runs are rejected against the tenant credit balance.

Grouped from the real @Controller('v1/…') set. See the OpenAPI spec below for the complete, authoritative operation list.

import { LokrixClient } from "@lokrix/sdk";

const lokrix = new LokrixClient({ apiKey: process.env.LOKRIX_API_KEY });

const { items } = await lokrix.projects.list();
const scan = await lokrix.scans.create({
  url: "https://acme.com",
  projectId: items[0].id,
});

Illustrative — the SDK is generated from the OpenAPI spec; confirm the exact export surface against the published @lokrix/sdk version. Python and Go SDKs are on the roadmap.

Is the API read-only on some plans?

On Starter, the REST API and MCP server are read-only. Growth and Enterprise unlock the full read/write API and MCP, plus BYOK (bring-your-own encrypted provider keys). The Free plan has no API/MCP access.

How do API-key scopes map to permissions?

Keys carry scopes from ApiKeyScopeSchema. Coarse legacy scopes (projects:read, scans:write, recommendations:read, usage:read) and fine-grained scopes (scores:read, results:read, analytics:read, alerts:read, alerts:write, webhooks:read, webhooks:write, …) both work: a route accepts either the coarse scope OR a matching fine-grained scope OR the admin superscope. Mint least-privilege keys per integration.

What is the difference between the stdio and HTTP MCP transports?

stdio (the default) runs lokrix-mcp as a subprocess of an MCP client such as Claude Desktop — best for a single local agent. Streamable HTTP (--http or MCP_HTTP_PORT) exposes the same tools over HTTP at /mcp — best for shared or remote agents. Both read LOKRIX_API_URL and LOKRIX_API_KEY from the environment.

Which tools and endpoints spend credits?

A credit is one live answer-engine response. Only live engine runs (start_scan / analyze against real engines) meter credits — prompts × engines × samples × cadence. Predictive scoring, what-if simulation (simulate_change), and path-to-#1 previews (get_optimization_plan) run on the calibrated model and cost 0 credits on every plan.

How do I read the OpenAPI spec and generate a client?

The API is NestJS + @nestjs/swagger. When docs are published, the raw OpenAPI 3.1 JSON is served at /openapi.json (with a /v1/openapi.json alias) and Swagger UI at /docs. Point any OpenAPI generator at the JSON to produce a typed client in your language, or use the maintained @lokrix/sdk (TypeScript).

What does every scored response carry?

Provenance (live / cached / predicted) and a 95% confidence interval on every probability. Responses are paginated with offset/limit ({ items, total, limit, offset }) and gzip/br compressed. Gated workspaces receive a 202 pending_approval envelope; over-budget requests return 429.