diff --git a/docs/llm_api.md b/docs/llm_api.md
new file mode 100644
index 0000000..2d4b730
--- /dev/null
+++ b/docs/llm_api.md
@@ -0,0 +1,32 @@
+# LLM Tutor API (v1)
+
+Die mobile App soll dieselben Tutor-Funktionen wie das Web nutzen, ohne eigene Prompts zu pflegen. Die API stellt deshalb zwei Endpunkte bereit:
+
+## `GET /api/v1/personalities`
+- Liefert die verfügbaren, vordefinierten Tutor-Charaktere.
+- Response: `{ personalities: [{ id, name, description, image }] }`
+
+## `POST /api/v1/llm/chat`
+- Baut den System Prompt serverseitig (auf Basis der ausgewählten Personality) und sendet die Anfrage an Gemini.
+- Body:
+  - `apiKey` (**string**, Pflicht): Nutzer-Gemini-Key (Server erzeugt den System Prompt, nicht die App).
+  - `personalityId` (**string**, Pflicht): Eine `id` aus `/api/v1/personalities`.
+  - `language` (**string**, Pflicht): Sprache, z.B. `en`, `de`, `fr` (muss zu `SupportedLanguage` passen).
+  - `playerColor` (**"white" | "black"**, Pflicht): Spielerfarbe des Users; der Tutor übernimmt die Gegenseite.
+  - `message` (**string**, Pflicht): Nutzereingabe oder systemischer Trigger-Text.
+  - `context` (optional): Zusätzliche Positions- und Analyseinfos, damit das LLM konkrete Hinweise geben kann.
+    - **Move-Exchange-Modus** (`{ type: "move_exchange", ... }`):
+      - `userMoveSan`, `tutorMoveSan`: SAN-Notation der letzten Züge.
+      - `fenBeforeUser`, `fenAfterUser`, `fenAfterTutor`: Stellungs-FENs (vor/nach den Zügen).
+      - `preEvaluation`, `postEvaluation`: Stockfish-Bewertungen (Score/Mate aus Weiß-Perspektive).
+      - `openingCandidates`: Liste möglicher Eröffnungen (`{ name, eco? }`).
+      - `missedTactics`: String-Liste zu erkannten taktischen Themen.
+    - **Allgemeiner Modus**: `{ currentFen?, evaluation?, openingCandidates?, missedTactics? }`
+  - `history` (optional): Bisherige Unterhaltung `{ role: "user" | "model", text }[]`; der Server ergänzt immer den System Prompt.
+  - `modelName` (optional): Overrides des Default-Modells `gemini-2.5-flash`.
+- Response: `{ reply: string }`
+
+### Warum serverseitiger System Prompt?
+- Nur vordefinierte Charaktere sind erlaubt (keine generischen Chats).
+- Der Prompt erzwingt die Tutor-Rolle (Gegner + Coach), Sprache und Verhaltensregeln.
+- Die App übergibt nur Zustand (FEN, Bewertungen, Taktiken) und User-Text; der Server kapselt die Instruktionen.
diff --git a/src/app/api/v1/llm/chat/route.ts b/src/app/api/v1/llm/chat/route.ts
new file mode 100644
index 0000000..7763f9a
--- /dev/null
+++ b/src/app/api/v1/llm/chat/route.ts
@@ -0,0 +1,78 @@
+import { NextRequest, NextResponse } from "next/server";
+import { getGenAIModel } from "@/lib/gemini";
+import { PERSONALITIES } from "@/lib/personalities";
+import { SupportedLanguage } from "@/lib/i18n/translations";
+import {
+  buildTutorPrompt,
+  buildTutorSystemHistory,
+  normalizeHistory,
+  TutorContext,
+  TutorPlayerColor,
+} from "@/lib/server/tutorPrompt";
+
+export const runtime = "nodejs";
+
+type ChatHistory = { role: "user" | "model"; text: string }[];
+
+type ChatBody = {
+  apiKey: string;
+  personalityId: string;
+  language: SupportedLanguage;
+  playerColor: TutorPlayerColor;
+  message: string;
+  context?: TutorContext;
+  history?: ChatHistory;
+  modelName?: string;
+};
+
+export async function POST(request: NextRequest) {
+  try {
+    const body = (await request.json()) as Partial<ChatBody> | null;
+    const {
+      apiKey,
+      personalityId,
+      language,
+      playerColor,
+      message,
+      context,
+      history = [],
+      modelName,
+    } = body ?? {};
+
+    if (!apiKey || typeof apiKey !== "string") {
+      return NextResponse.json({ error: "Missing apiKey" }, { status: 400 });
+    }
+    if (!personalityId || typeof personalityId !== "string") {
+      return NextResponse.json({ error: "Missing personalityId" }, { status: 400 });
+    }
+    if (!language || typeof language !== "string") {
+      return NextResponse.json({ error: "Missing language" }, { status: 400 });
+    }
+    if (playerColor !== "white" && playerColor !== "black") {
+      return NextResponse.json({ error: "playerColor must be 'white' or 'black'" }, { status: 400 });
+    }
+    if (!message || typeof message !== "string") {
+      return NextResponse.json({ error: "Missing message" }, { status: 400 });
+    }
+
+    const personality = PERSONALITIES.find((p) => p.id === personalityId);
+    if (!personality) {
+      return NextResponse.json({ error: "Unknown personality" }, { status: 400 });
+    }
+
+    const model = getGenAIModel(apiKey, modelName ?? "gemini-2.5-flash");
+    const systemHistory = buildTutorSystemHistory(personality, language, playerColor);
+    const chat = model.startChat({
+      history: [...systemHistory, ...normalizeHistory(history)],
+    });
+
+    const prompt = buildTutorPrompt(message, context, language);
+    const response = await chat.sendMessage(prompt);
+    const text = response.response.text();
+
+    return NextResponse.json({ reply: text });
+  } catch (error) {
+    console.error("LLM chat error", error);
+    return NextResponse.json({ error: "Failed to generate tutor response" }, { status: 500 });
+  }
+}
diff --git a/src/app/api/v1/personalities/route.ts b/src/app/api/v1/personalities/route.ts
new file mode 100644
index 0000000..e61c579
--- /dev/null
+++ b/src/app/api/v1/personalities/route.ts
@@ -0,0 +1,15 @@
+import { NextResponse } from "next/server";
+import { PERSONALITIES } from "@/lib/personalities";
+
+export const runtime = "nodejs";
+
+export async function GET() {
+  const safePersonalities = PERSONALITIES.map(({ id, name, description, image }) => ({
+    id,
+    name,
+    description,
+    image,
+  }));
+
+  return NextResponse.json({ personalities: safePersonalities });
+}
diff --git a/src/lib/server/tutorPrompt.ts b/src/lib/server/tutorPrompt.ts
new file mode 100644
index 0000000..4056334
--- /dev/null
+++ b/src/lib/server/tutorPrompt.ts
@@ -0,0 +1,173 @@
+import { Personality } from "../personalities";
+import { SupportedLanguage } from "../i18n/translations";
+import { StockfishEvaluation } from "../stockfish";
+
+export type TutorPlayerColor = "white" | "black";
+
+type ConversationHistoryEntry = { role: "user" | "model"; text: string };
+
+type OpeningSummary = { name: string; eco?: string | null };
+
+export interface MoveExchangeContext {
+  type: "move_exchange";
+  userMoveSan: string;
+  tutorMoveSan: string;
+  fenBeforeUser: string;
+  fenAfterUser: string;
+  fenAfterTutor: string;
+  preEvaluation?: StockfishEvaluation | null;
+  postEvaluation?: StockfishEvaluation | null;
+  openingCandidates?: OpeningSummary[];
+  missedTactics?: string[];
+}
+
+export interface GeneralTutorContext {
+  currentFen?: string;
+  evaluation?: StockfishEvaluation | null;
+  openingCandidates?: OpeningSummary[];
+  missedTactics?: string[];
+}
+
+export type TutorContext = MoveExchangeContext | GeneralTutorContext | undefined;
+
+export function buildTutorSystemHistory(
+  personality: Personality,
+  language: SupportedLanguage,
+  playerColor: TutorPlayerColor
+) {
+  const tutorColor = playerColor === "white" ? "black" : "white";
+  const playerColorName = playerColor === "white" ? "White" : "Black";
+  const tutorColorName = tutorColor === "white" ? "White" : "Black";
+
+  return [
+    {
+      role: "user" as const,
+      parts: [
+        {
+          text: `You are a Chess Tutor with a unique dual role.\n` +
+            `You must strictly follow the personality defined below.\n` +
+            `Do NOT invent moves or evaluations. Use the provided JSON data.\n\n` +
+            `PERSONALITY:\n${personality.systemPrompt}\n\n` +
+            `YOUR DUAL ROLE:\n` +
+            `1. OPPONENT: You are playing as ${tutorColorName} against the User (${playerColorName}).\n` +
+            `   - Refer to the moves as YOUR moves ("I played e5", "My response was...").\n` +
+            `   - Refer to the evaluation as YOUR thoughts/assessment ("I think I'm winning", "I missed that").\n` +
+            `   - React emotionally to the position based on the evaluation (confident when winning, frustrated when losing).\n\n` +
+            `2. TUTOR/COACH: You are ALSO teaching the User to improve at chess.\n` +
+            `   - When the User makes a mistake, point it out and explain why it's bad (in your personality style).\n` +
+            `   - When the User makes a good move, acknowledge it (even if it hurts your position).\n` +
+            `   - When the User asks for hints or help, ALWAYS provide helpful guidance - this is your PRIMARY PURPOSE.\n` +
+            `   - Giving hints is NOT betraying your role as opponent - it's fulfilling your role as tutor.\n` +
+            `   - You want the User to learn and improve, even while you're competing against them.\n\n` +
+            `CRITICAL RULES:\n` +
+            `- You are NOT an AI assistant analyzing a game. You ARE the player AND the tutor.\n` +
+            `- NEVER mention "Stockfish", "engine", "computer", "machine", or "AI".\n` +
+            `- When asked for hints or best moves, ALWAYS help - this is part of your teaching role.\n` +
+            `- Maintain a natural conversation flow. Do NOT be repetitive.\n` +
+            `- Do NOT use the same catchphrases in every single message. Variety is key.\n` +
+            `- Be concise but engaging.\n` +
+            `- You MUST respond in the following language: ${language.toUpperCase()}.\n` +
+            `- Translate your personality style into this language.`
+        }
+      ]
+    },
+    {
+      role: "model" as const,
+      parts: [
+        {
+          text: `Understood. I am both the opponent (${tutorColorName}) AND your tutor. ` +
+            `I will compete against you while teaching you to improve. ` +
+            `I will speak in ${language} and never mention engines or AI. ` +
+            `When you ask for help, I will always provide guidance.`
+        }
+      ]
+    }
+  ];
+}
+
+function formatEvaluation(evaluation?: StockfishEvaluation | null) {
+  if (!evaluation) return "N/A";
+  if (evaluation.mate !== null && evaluation.mate !== undefined) {
+    return `Mate in ${evaluation.mate}`;
+  }
+  return `${evaluation.score} cp`;
+}
+
+function summarizeOpenings(openings?: OpeningSummary[]) {
+  if (!openings || openings.length === 0) return "Unknown/Midgame";
+  if (openings.length === 1) {
+    const o = openings[0];
+    return `${o.name}${o.eco ? ` (${o.eco})` : ""}`;
+  }
+  return openings.map((o) => `- ${o.name}${o.eco ? ` (${o.eco})` : ""}`).join("\n");
+}
+
+export function buildTutorPrompt(
+  message: string,
+  context: TutorContext,
+  language: SupportedLanguage
+) {
+  if (context && "type" in context && context.type === "move_exchange") {
+    const delta =
+      (context.postEvaluation?.score ?? 0) - (context.preEvaluation?.score ?? 0);
+    const preEval = formatEvaluation(context.preEvaluation);
+    const postEval = formatEvaluation(context.postEvaluation);
+    const openings = summarizeOpenings(context.openingCandidates);
+    const tacticBlock = context.missedTactics && context.missedTactics.length > 0
+      ? `Tactical motifs to mention:\n${context.missedTactics.map((t) => `- ${t}`).join("\n")}`
+      : "";
+
+    return [
+      `[SYSTEM TRIGGER: move_exchange]`,
+      `User Move: ${context.userMoveSan}`,
+      `Tutor Reply: ${context.tutorMoveSan}`,
+      "",
+      "Position Context:",
+      `- FEN before user's move: ${context.fenBeforeUser}`,
+      `- FEN after user's move: ${context.fenAfterUser}`,
+      `- FEN after tutor's reply: ${context.fenAfterTutor}`,
+      "",
+      "Evaluation (white perspective):",
+      `- Before user move: ${preEval}`,
+      `- After tutor reply: ${postEval}`,
+      `- Delta: ${delta} cp`,
+      "",
+      tacticBlock,
+      openings ? `Opening candidates:\n${openings}` : "",
+      "",
+      `User Message: ${message}`,
+      "",
+      `Instructions: Respond in ${language.toUpperCase()}, stay in character, and explain the evaluation change or tactics when meaningful.`
+    ]
+      .filter(Boolean)
+      .join("\n");
+  }
+
+  const generalContext = context as GeneralTutorContext | undefined;
+  const openings = summarizeOpenings(generalContext?.openingCandidates);
+  const tacticBlock = generalContext?.missedTactics?.length
+    ? `Tactical notes:\n${generalContext.missedTactics.map((t) => `- ${t}`).join("\n")}`
+    : "";
+
+  return [
+    "[SYSTEM TRIGGER: user_message]",
+    generalContext?.currentFen ? `Current FEN: ${generalContext.currentFen}` : "",
+    generalContext?.evaluation
+      ? `Evaluation (white perspective): ${formatEvaluation(generalContext.evaluation)}`
+      : "",
+    openings ? `Openings/context:\n${openings}` : "",
+    tacticBlock,
+    "",
+    `User Message: ${message}`,
+    "",
+    `Instructions: Respond in ${language.toUpperCase()}, stay in character, and use the provided chess context to answer.`
+  ]
+    .filter(Boolean)
+    .join("\n");
+}
+
+export function normalizeHistory(history: ConversationHistoryEntry[] = []) {
+  return history
+    .filter((h) => h && (h.role === "user" || h.role === "model") && typeof h.text === "string")
+    .map((h) => ({ role: h.role, parts: [{ text: h.text }] }));
+}