feat(plugins): add llama.cpp server integration and widget (#6765)

apps/docs/docs/integrations/llama-cpp/index.mdx

@@ -0,0 +1,55 @@
+---
+title: "llama.cpp"
+description: "Connect Homarr to a llama.cpp llama-server to monitor your local LLM: health, loaded model, generation speed and request status."
+hide_title: true
+---
+
+import { IntegrationHeader } from "@site/src/components/integrations/header";
+import { IntegrationCapabilites } from "@site/src/components/integrations/widgets";
+import { AddingIntegration } from "@site/src/components/integrations/adding";
+import { IntegrationSecrets } from "@site/src/components/integrations/secrets";
+import { llamacppIntegration } from ".";
+import { llamacppWidget } from "@site/docs/widgets/llama-cpp";
+
+<IntegrationHeader integration={llamacppIntegration} categories={["System Monitoring"]} />
+
+The llama.cpp integration lets Homarr connect to a [llama-server](https://github.com/ggml-org/llama.cpp) instance so you can see the health of your local LLM, which model is loaded, and how fast it is generating — all from your dashboard.
+
+### Widgets & Capabilities
+
+<IntegrationCapabilites items={[{ widget: llamacppWidget }]} />
+
+### Adding the integration
+
+<AddingIntegration />
+
+:::info URL configuration
+Use the root URL of your `llama-server`, for example `http://192.168.1.50:8080`. Homarr calls the `/health`, `/v1/models`, `/metrics` and `/slots` endpoints under that URL automatically.
+:::
+
+:::tip Metrics
+Generation speed and token counters are read from the Prometheus endpoint at `/metrics`. Start `llama-server` with the `--metrics` flag so these values are available:
+
+```shell
+llama-server --model /path/to/model.gguf --metrics
+```
+
+Without `--metrics` the widget still shows health, model information, and active-request speed from `/slots`; metrics-backed counters and aggregate speed stay empty.
+:::
+
+:::info Context and per-request speed
+Context (KV cache) usage and the in-flight request data are read from the `/slots` endpoint, which is available on every recent llama-server build and requires no extra flags. The per-request generation speed shown while a request is running is derived by tracking the request's decoded token count across polls, so it reflects the average speed of the request that is currently being generated.
+:::
+
+### Secrets
+
+llama-server does not require authentication by default. The integration only needs the URL of your server.
+
+<IntegrationSecrets
+  secrets={[
+    {
+      credentials: [],
+      steps: [],
+    },
+  ]}
+/>

apps/docs/docs/integrations/llama-cpp/index.ts

@@ -0,0 +1,9 @@
+import { IntegrationDefinition } from "@site/src/types";
+
+export const llamacppIntegration: IntegrationDefinition = {
+  name: "llama.cpp",
+  description:
+    "Connect to a local llama.cpp llama-server to show model, throughput and request status on your dashboard.",
+  iconUrl: "https://cdn.jsdelivr.net/gh/selfhst/icons/svg/llama-cpp.svg",
+  path: "../../integrations/llama-cpp",
+};

apps/docs/docs/widgets/llama-cpp/index.mdx

@@ -0,0 +1,43 @@
+---
+title: "llama.cpp"
+description: "Displays the health, loaded model and generation speed of a local llama.cpp llama-server."
+hide_title: true
+---
+
+import { WidgetHeader } from "@site/src/components/widgets/header";
+import { AddingWidget } from "@site/src/components/widgets/adding";
+import { WidgetConfig } from "@site/src/components/widgets/configuration";
+import { WidgetIntegrations } from "@site/src/components/widgets/integrations";
+import { llamacppWidget } from ".";
+import { llamacppIntegration } from "@site/docs/integrations/llama-cpp";
+
+<WidgetHeader widget={llamacppWidget} categories={["System Monitoring"]} />
+
+This widget shows the live status of your local [llama.cpp](https://github.com/ggml-org/llama.cpp) `llama-server`: whether it is online, which model is loaded (with quantization, context size, file size and parameter count), the generation speed in tokens per second, and how many requests are being processed. When the server is started with `--metrics`, it also reports total token counts and the prompt cache hit rate.
+
+### Generation speed
+
+The widget picks the most meaningful speed for the current state (hover the speed for an explanation of which one is shown):
+
+- **While a request is being generated** — the average speed of that request since it started, derived from the token count reported by `/slots`.
+- **While the server is busy but the per-request data is unavailable** — the live speed from the `/metrics` rate gauge.
+- **While the server is idle** — the average generation speed since the server started (cumulative tokens ÷ seconds), so the value stays meaningful instead of dropping to zero.
+
+### Supported Integrations
+
+<WidgetIntegrations
+  items={[
+    {
+      integration: llamacppIntegration,
+      note: "Displays health, model details, generation speed and request status",
+    },
+  ]}
+/>
+
+### Adding the widget
+
+<AddingWidget />
+
+### Configuration
+
+<WidgetConfig configuration={llamacppWidget.configuration} />

apps/docs/docs/widgets/llama-cpp/index.ts

@@ -0,0 +1,37 @@
+import { WidgetDefinition } from "@site/src/types";
+import { IconCpu } from "@tabler/icons-react";
+
+export const llamacppWidget: WidgetDefinition = {
+  icon: IconCpu,
+  name: "llama.cpp",
+  description: "Shows the health, loaded model and generation speed of a local llama.cpp llama-server.",
+  path: "../../widgets/llama-cpp",
+  configuration: {
+    items: [
+      {
+        name: "Show title",
+        description: "Displays the llama.cpp label and status badge at the top of the widget",
+        values: { type: "boolean" },
+        defaultValue: "yes",
+      },
+      {
+        name: "Show model info",
+        description: "Displays the loaded model name, quantization, context size and file size",
+        values: { type: "boolean" },
+        defaultValue: "yes",
+      },
+      {
+        name: "Show cache hit rate",
+        description: "Displays the prompt cache hit rate as a percentage bar",
+        values: { type: "boolean" },
+        defaultValue: "yes",
+      },
+      {
+        name: "Show context usage",
+        description: "Displays the context (KV cache) token usage as a progress bar",
+        values: { type: "boolean" },
+        defaultValue: "yes",
+      },
+    ],
+  },
+};

packages/api/src/router/widgets/index.ts

@@ -46,4 +46,5 @@ export const widgetRouter = createTRPCRouter({
   customApi: lazy(() => import("./custom-api").then((mod) => mod.customApiRouter)),
   secrets: lazy(() => import("./widget-secrets").then((mod) => mod.widgetSecretsRouter)),
   wud: lazy(() => import("./wud").then((mod) => mod.wudRouter)),
+  llamacpp: lazy(() => import("./llamacpp").then((mod) => mod.llamacppRouter)),
 });

packages/api/src/router/widgets/llamacpp.ts

@@ -0,0 +1,29 @@
+import { llamacppStatsRequestHandler } from "@homarr/request-handler/llamacpp";
+import { mockWidgetData } from "@homarr/integrations";
+
+import { createOneWidgetIntegrationMiddleware } from "../../middlewares/integration";
+import { createTRPCRouter, publicProcedure } from "../../trpc";
+
+export const llamacppRouter = createTRPCRouter({
+  getStats: publicProcedure
+    .meta({
+      mcp: {
+        enabled: true,
+        description:
+          "Returns health, loaded model and Prometheus metrics (generation speed, token counts, request queue) from a llama.cpp (llama-server) integration. REQUIRED: integrationId from integration_all. The caller needs query permission for that integration.",
+      },
+    })
+    .concat(createOneWidgetIntegrationMiddleware("query", "llamacpp"))
+    .query(async ({ ctx }) => {
+      if (ctx.integration.kind === "mock") {
+        return { stats: mockWidgetData.llamacpp, updatedAt: new Date(mockWidgetData.timestamp) };
+      }
+      const handler = llamacppStatsRequestHandler.handler({ ...ctx.integration, kind: "llamacpp" }, {});
+      const { data, timestamp } = await handler.getDataAsync();
+
+      return {
+        stats: data,
+        updatedAt: timestamp,
+      };
+    }),
+});

packages/db/schema/mysql.ts

@@ -505,6 +505,7 @@ export const assistantConfigurations = mysqlTable("assistant_configuration", {
       | "together"
       | "ollama"
       | "lm-studio"
+      | "llama-cpp"
       | "custom"
     >()
     .notNull()

packages/db/schema/postgresql.ts

@@ -503,6 +503,7 @@ export const assistantConfigurations = pgTable("assistant_configuration", {
       | "together"
       | "ollama"
       | "lm-studio"
+      | "llama-cpp"
       | "custom"
     >()
     .notNull()

packages/db/schema/sqlite.ts

@@ -488,6 +488,7 @@ export const assistantConfigurations = sqliteTable("assistant_configuration", {
       | "together"
       | "ollama"
       | "lm-studio"
+      | "llama-cpp"
       | "custom"
     >()
     .notNull()

packages/definitions/src/assistant.ts

@@ -11,6 +11,7 @@ export const assistantProviderIds = [
   "together",
   "ollama",
   "lm-studio",
+  "llama-cpp",
   "custom",
 ] as const;
 
@@ -156,6 +157,15 @@ export const assistantProviderPresets = {
     discoveryAuthentication: "bearer",
     iconUrl: `${providerIconBaseUrl}/lmstudio.svg`,
   },
+  "llama-cpp": {
+    label: "llama.cpp",
+    baseUrl: "http://localhost:8080/v1",
+    modelDiscoveryPath: "/models",
+    requiresApiKey: false,
+    category: "local",
+    discoveryAuthentication: "bearer",
+    iconUrl: "https://cdn.jsdelivr.net/gh/selfhst/icons/svg/llama-cpp.svg",
+  },
   custom: {
     label: undefined,
     baseUrl: "",

packages/definitions/src/docs/homarr-docs-sitemap.ts

@@ -333,6 +333,7 @@ export type HomarrDocumentationPath =
   | "/docs/integrations/traefik"
   | "/docs/integrations/archiveteam-warrior"
   | "/docs/integrations/whats-up-docker"
+  | "/docs/integrations/llama-cpp"
   | "/docs/widgets/clock"
   | "/docs/widgets/weather"
   | "/docs/widgets/air-quality"
@@ -390,4 +391,5 @@ export type HomarrDocumentationPath =
   | "/docs/widgets/traefik"
   | "/docs/widgets/custom-api"
   | "/docs/widgets/assistant"
-  | "/docs/widgets/whats-up-docker";
+  | "/docs/widgets/whats-up-docker"
+  | "/docs/widgets/llama-cpp";

packages/definitions/src/docs/widget-doc-slugs.ts

@@ -60,4 +60,5 @@ export const widgetDocSlugs = {
   customApi: "custom-api",
   assistant: "assistant",
   wud: "whats-up-docker",
+  llamacpp: "llama-cpp",
 } satisfies Record<WidgetKind, string | null>;

packages/definitions/src/integration.ts

@@ -545,6 +545,14 @@ export const integrationDefs = {
     documentationSlug: "whats-up-docker",
     defaultPort: 3000,
   },
+  llamacpp: {
+    name: "llama.cpp",
+    secretKinds: [[]],
+    iconUrl: "https://cdn.jsdelivr.net/gh/selfhst/icons/svg/llama-cpp.svg",
+    category: ["systemMonitoring"],
+    documentationSlug: "llama-cpp",
+    defaultPort: 8080,
+  },
   // This integration only returns mock data, it is used during development (but can also be used in production by directly going to the create page)
   mock: {
     name: "Mock",
@@ -713,6 +721,7 @@ export const integrationCategories = [
   "uptimeMonitoring",
   "subtitleManager",
   "reverseProxy",
+  "systemMonitoring",
 ] as const;
 
 export type IntegrationCategory = (typeof integrationCategories)[number];

packages/definitions/src/test/widget-integration-map.spec.ts

@@ -48,6 +48,7 @@ describe("widget integration config", () => {
       archiveTeamWarrior: 1,
       anchorNote: 1,
       wud: 1,
+      llamacpp: 1,
     });
   });
 });

packages/definitions/src/widget-integration-map.ts

@@ -87,6 +87,7 @@ export const widgetIntegrationConfigs = {
   anchorNote: { supportedIntegrations: ["anchor", "mock"], maxIntegrations: 1 },
   traefik: { supportedIntegrations: ["traefik", "mock"] },
   wud: { supportedIntegrations: ["wud", "mock"], maxIntegrations: 1 },
+  llamacpp: { supportedIntegrations: ["llamacpp", "mock"], maxIntegrations: 1 },
 } satisfies Partial<Record<WidgetKind, WidgetIntegrationConfig>>;
 
 export type WidgetKindWithIntegration = keyof typeof widgetIntegrationConfigs;

packages/definitions/src/widget.ts

@@ -58,6 +58,7 @@ export const widgetKinds = [
   "customApi",
   "assistant",
   "wud",
+  "llamacpp",
 ] as const;
 
 export type WidgetKind = (typeof widgetKinds)[number];

packages/integrations/src/base/creator.ts

@@ -110,6 +110,8 @@ const integrationCreators = {
   traefik: async (input: IntegrationInput) =>
     new (await import("../traefik/traefik-integration")).TraefikIntegration(input),
   wud: async (input: IntegrationInput) => new (await import("../wud/wud-integration")).WudIntegration(input),
+  llamacpp: async (input: IntegrationInput) =>
+    new (await import("../llama-cpp/llamacpp-integration")).LlamacppIntegration(input),
 } satisfies Record<IntegrationKind, (input: IntegrationInput) => Promise<Integration>>;
 
 type IntegrationCreators = typeof integrationCreators;

packages/integrations/src/index.ts

@@ -1,6 +1,7 @@
 // General integrations
 export { ArchiveTeamWarriorIntegration } from "./archive-team-warrior/archive-team-warrior-integration";
 export { WudIntegration } from "./wud/wud-integration";
+export { LlamacppIntegration } from "./llama-cpp/llamacpp-integration";
 export { AdGuardHomeIntegration } from "./adguard-home/adguard-home-integration";
 export { TechnitiumDnsIntegration } from "./technitium/technitium-integration";
 export { AnchorIntegration } from "./anchor/anchor-integration";
@@ -95,6 +96,7 @@ export type {
   ArchiveTeamWarriorStatus,
 } from "./archive-team-warrior/archive-team-warrior-types";
 export type { WudStats, WudContainerUpdate } from "./wud/wud-types";
+export type { LlamacppStats, LlamacppModel } from "./llama-cpp/llamacpp-types";
 
 // Schemas
 export { anchorNotesListInputSchema, anchorNoteUpdateInputSchema } from "./anchor/anchor-types";

packages/integrations/src/llama-cpp/llamacpp-integration.ts

@@ -0,0 +1,83 @@
+import { ParseError, ResponseError } from "@homarr/common/server";
+import { fetchWithTrustedCertificatesAsync } from "@homarr/core/infrastructure/http";
+
+import type { IntegrationTestingInput } from "../base/integration";
+import { Integration } from "../base/integration";
+import { TestConnectionError } from "../base/test-connection/test-connection-error";
+import type { TestingResult } from "../base/test-connection/test-connection-service";
+
+import type { LlamacppStats } from "./llamacpp-types";
+import {
+  mapContextUsage,
+  mapLlamacppModel,
+  mapLlamacppPerRequest,
+  mapLlamacppStats,
+  mapSlotsSummary,
+  parseLlamacppHealthAsync,
+  parseLlamacppMetricsAsync,
+  parseLlamacppModelsAsync,
+  parseLlamacppSlotsAsync,
+} from "./llamacpp-types";
+
+const LLMACPP_REQUEST_TIMEOUT_MS = 10_000;
+
+export class LlamacppIntegration extends Integration {
+  protected async testingAsync(input: IntegrationTestingInput): Promise<TestingResult> {
+    const response = await input.fetchAsync(this.url("/health"), {
+      signal: AbortSignal.timeout(LLMACPP_REQUEST_TIMEOUT_MS),
+    });
+
+    if (!response.ok) {
+      return TestConnectionError.StatusResult(response);
+    }
+
+    // Validate the health payload shape, not just that it is parseable JSON: an
+    // unrelated JSON body must not pass connection testing.
+    try {
+      await parseLlamacppHealthAsync(response);
+    } catch (error) {
+      if (error instanceof ParseError) {
+        return TestConnectionError.ParseResult(error);
+      }
+      return TestConnectionError.UnknownResult(error);
+    }
+
+    return { success: true };
+  }
+
+  public async getStatsAsync(): Promise<LlamacppStats> {
+    const [healthResponse, modelsResponse, metricsResponse, slotsResponse] = await Promise.all([
+      fetchWithTrustedCertificatesAsync(this.url("/health"), { timeout: LLMACPP_REQUEST_TIMEOUT_MS }),
+      fetchWithTrustedCertificatesAsync(this.url("/v1/models"), { timeout: LLMACPP_REQUEST_TIMEOUT_MS }),
+      fetchWithTrustedCertificatesAsync(this.url("/metrics"), { timeout: LLMACPP_REQUEST_TIMEOUT_MS }),
+      // /slots is not present on every build; a missing endpoint degrades to no context usage.
+      fetchWithTrustedCertificatesAsync(this.url("/slots"), { timeout: LLMACPP_REQUEST_TIMEOUT_MS }).catch(() => null),
+    ]);
+
+    if (!healthResponse.ok) {
+      throw new ResponseError(healthResponse);
+    }
+    if (!modelsResponse.ok) {
+      throw new ResponseError(modelsResponse);
+    }
+
+    // /metrics only exists when llama-server was started with --metrics; a non-OK
+    // response (typically 501 Not Implemented) degrades to no rate counters instead
+    // of dropping the health and model data that the other endpoints still provide.
+    const metricsAvailable = metricsResponse.ok;
+
+    const [health, models, metrics, slots] = await Promise.all([
+      parseLlamacppHealthAsync(healthResponse),
+      parseLlamacppModelsAsync(modelsResponse),
+      metricsAvailable ? parseLlamacppMetricsAsync(metricsResponse) : Promise.resolve([]),
+      slotsResponse?.ok ? parseLlamacppSlotsAsync(slotsResponse) : Promise.resolve(null),
+    ]);
+
+    const firstModel = models.data[0] ? mapLlamacppModel(models.data[0]) : null;
+    const contextUsage = mapContextUsage(slots);
+    const slotsSummary = mapSlotsSummary(slots);
+    const perRequest = mapLlamacppPerRequest(slots);
+
+    return mapLlamacppStats(health, firstModel, metrics, contextUsage, slotsSummary, perRequest);
+  }
+}

packages/integrations/src/llama-cpp/llamacpp-types.ts

@@ -0,0 +1,381 @@
+import { ParseError } from "@homarr/common/server";
+import { z } from "zod/v4";
+
+export const LLMACPP_HEALTH_PARSE_ERROR_MESSAGE = "Invalid llama.cpp /health response";
+export const LLMACPP_MODELS_PARSE_ERROR_MESSAGE = "Invalid llama.cpp /v1/models response";
+
+export const parsePrometheusMetrics = (text: string): { name: string; value: number }[] => {
+  const metrics: { name: string; value: number }[] = [];
+  for (const rawLine of text.split(/\r?\n/)) {
+    const line = rawLine.trim();
+    if (line.length === 0 || line.startsWith("#")) {
+      continue;
+    }
+
+    const match = /^([a-zA-Z_:][a-zA-Z0-9_:]*)(?:\{[^}]*\})?\s+(-?[0-9]+(?:\.[0-9]+)?(?:[eE][+-]?[0-9]+)?|NaN)$/.exec(
+      line,
+    );
+    if (!match) {
+      continue;
+    }
+
+    const name = match[1];
+    const rawValue = match[2];
+    if (!name || rawValue === undefined) {
+      continue;
+    }
+
+    const value = Number.parseFloat(rawValue);
+    if (!Number.isFinite(value)) {
+      continue;
+    }
+
+    metrics.push({ name, value });
+  }
+
+  return metrics;
+};
+
+export const llamacppHealthSchema = z.object({
+  status: z.string(),
+});
+
+const llamacppModelMetaSchema = z
+  .object({
+    n_ctx: z.number().optional(),
+    n_params: z.number().optional(),
+    size: z.number().optional(),
+    ftype: z.string().optional(),
+  })
+  .optional();
+
+const llamacppModelSchema = z.object({
+  id: z.string(),
+  meta: llamacppModelMetaSchema,
+});
+
+export const llamacppModelsSchema = z.object({
+  data: z.array(llamacppModelSchema),
+});
+
+export interface LlamacppModel {
+  id: string;
+  name: string;
+  contextSize: number | null;
+  parameterCount: number | null;
+  fileSizeBytes: number | null;
+  quantization: string | null;
+}
+
+export interface LlamacppPerRequest {
+  taskId: number | null;
+  decodedTokens: number | null;
+}
+
+export interface LlamacppSlotsSummary {
+  total: number;
+  processing: number;
+  decoding: number;
+  speculative: boolean | null;
+}
+
+export interface LlamacppStats {
+  health: string;
+  model: LlamacppModel | null;
+  contextUsage: LlamacppContextUsage | null;
+  slots: LlamacppSlotsSummary | null;
+  metrics: {
+    generationSpeedTps: number | null;
+    promptSpeedTps: number | null;
+    avgGenerationSpeedTps: number | null;
+    avgPromptSpeedTps: number | null;
+    requestsProcessing: number | null;
+    requestsDeferred: number | null;
+    tokensProcessed: number | null;
+    tokensGenerated: number | null;
+    promptCacheHitRate: number | null;
+    requestDecodedTokens: number | null;
+    taskId: number | null;
+    specDraftTokens: number | null;
+    specAcceptedTokens: number | null;
+    specDrafts: number | null;
+  };
+}
+
+const shortModelName = (id: string): string =>
+  id
+    .split("/")
+    .pop()
+    ?.replace(/\.gguf$/i, "") ?? id;
+
+export const mapLlamacppModel = (model: z.infer<typeof llamacppModelSchema>): LlamacppModel => ({
+  id: model.id,
+  name: shortModelName(model.id),
+  contextSize: model.meta?.n_ctx ?? null,
+  parameterCount: model.meta?.n_params ?? null,
+  fileSizeBytes: model.meta?.size ?? null,
+  quantization: model.meta?.ftype ?? null,
+});
+
+export const getMetricValue = (metrics: readonly { name: string; value: number }[], name: string): number | null =>
+  metrics.find((metric) => metric.name === name)?.value ?? null;
+
+/**
+ * Averages cumulative counters (tokens ÷ seconds) so a stable number is available while
+ * the server is idle, mirroring how llama-monitor derives its always-on speeds.
+ */
+export const avgTokensPerSecond = (tokens: number | null, seconds: number | null): number | null =>
+  tokens !== null && seconds !== null && seconds > 0 ? tokens / seconds : null;
+
+interface LlamacppSlot {
+  n_ctx?: number;
+  n_prompt_tokens?: number;
+  id_task?: number;
+  is_processing?: boolean;
+  speculative?: boolean;
+  next_token?: { n_decoded?: number }[];
+}
+
+export interface LlamacppContextUsage {
+  usedTokens: number;
+  contextSize: number;
+  percent: number;
+}
+
+const clampPercent = (value: number): number => Math.min(100, Math.max(0, value));
+
+/**
+ * Computes context (KV) usage from the /slots endpoint. The /metrics endpoint has no
+ * KV metrics, so the largest used prompt tokens over the slot context size is used.
+ */
+export const mapContextUsage = (slots: readonly unknown[] | null): LlamacppContextUsage | null => {
+  if (!Array.isArray(slots)) {
+    return null;
+  }
+
+  let largest: LlamacppContextUsage | null = null;
+  for (const slot of slots) {
+    if (typeof slot !== "object" || slot === null) {
+      continue;
+    }
+
+    const record = slot as Partial<LlamacppSlot>;
+    const contextSize = record.n_ctx;
+    const usedTokens = record.n_prompt_tokens;
+
+    if (typeof contextSize === "number" && contextSize > 0 && typeof usedTokens === "number" && usedTokens >= 0) {
+      const usage: LlamacppContextUsage = {
+        usedTokens,
+        contextSize,
+        percent: clampPercent(Math.round((usedTokens / contextSize) * 1000) / 10),
+      };
+      if (largest === null || usage.usedTokens > largest.usedTokens) {
+        largest = usage;
+      }
+    }
+  }
+
+  return largest;
+};
+
+/**
+ * Summarizes the /slots endpoint: total slot count, how many are currently
+ * processing, how many are actively decoding, and whether any slot has
+ * speculative decoding enabled. Null when /slots is unavailable.
+ */
+export const mapSlotsSummary = (slots: readonly unknown[] | null): LlamacppSlotsSummary | null => {
+  if (!Array.isArray(slots) || slots.length === 0) {
+    return null;
+  }
+
+  let total = 0;
+  let processing = 0;
+  let decoding = 0;
+  let speculative: boolean | null = null;
+
+  for (const slot of slots) {
+    if (typeof slot !== "object" || slot === null) {
+      continue;
+    }
+
+    total += 1;
+
+    const record = slot as Partial<LlamacppSlot>;
+    if (record.is_processing === true) {
+      processing += 1;
+      if (Array.isArray(record.next_token) && record.next_token.length > 0) {
+        decoding += 1;
+      }
+    }
+    if (record.speculative === true) {
+      speculative = true;
+    } else if (typeof record.speculative === "boolean" && speculative !== true) {
+      speculative = record.speculative;
+    }
+  }
+
+  return total > 0
+    ? {
+        total,
+        processing,
+        decoding,
+        speculative,
+      }
+    : null;
+};
+
+/**
+ * Extracts the active request's identity and generated-token count from /slots. `id_task`
+ * identifies the in-flight request (llama.cpp task IDs are monotonically increasing and
+ * never reused) and `next_token[0].n_decoded` is `stats.n_gen`, the tokens this request has
+ * generated so far, which the server resets to 0 whenever a new generation starts. The widget
+ * takes deltas between polls to derive the per-request average speed.
+ */
+export const mapLlamacppPerRequest = (slots: readonly unknown[] | null): LlamacppPerRequest => {
+  if (!Array.isArray(slots)) {
+    return { taskId: null, decodedTokens: null };
+  }
+
+  for (const slot of slots) {
+    if (typeof slot !== "object" || slot === null) {
+      continue;
+    }
+
+    const record = slot as LlamacppSlot;
+    if (record.is_processing !== true) {
+      continue;
+    }
+
+    const taskId = typeof record.id_task === "number" ? record.id_task : null;
+    const nextToken = record.next_token?.[0];
+    const decodedTokens = typeof nextToken?.n_decoded === "number" ? nextToken.n_decoded : null;
+
+    return { taskId, decodedTokens };
+  }
+
+  return { taskId: null, decodedTokens: null };
+};
+
+export const requestSpeedTps = (tokens: number | null, elapsedSeconds: number | null): number | null => {
+  if (tokens === null || elapsedSeconds === null || elapsedSeconds <= 0) {
+    return null;
+  }
+  const speed = tokens / elapsedSeconds;
+  return Number.isFinite(speed) && speed > 0 ? speed : null;
+};
+
+export const parseLlamacppSlotsAsync = async (response: {
+  json: () => Promise<unknown>;
+}): Promise<unknown[] | null> => {
+  let json: unknown;
+  try {
+    json = await response.json();
+  } catch (error) {
+    throw new ParseError("Invalid llama.cpp /slots response", {
+      cause: error instanceof Error ? error : new Error(String(error)),
+    });
+  }
+
+  return Array.isArray(json) ? json : null;
+};
+
+export const mapLlamacppStats = (
+  health: z.infer<typeof llamacppHealthSchema>,
+  model: LlamacppModel | null,
+  metrics: readonly { name: string; value: number }[],
+  contextUsage: LlamacppContextUsage | null,
+  slotsSummary: LlamacppSlotsSummary | null,
+  perRequest: LlamacppPerRequest,
+): LlamacppStats => {
+  const promptTokens = getMetricValue(metrics, "llamacpp:prompt_tokens_total");
+  const cachedTokens = getMetricValue(metrics, "llamacpp:prompt_tokens_cached_total");
+
+  const promptCacheHitRate =
+    promptTokens !== null && cachedTokens !== null && promptTokens + cachedTokens > 0
+      ? Math.round((cachedTokens / (promptTokens + cachedTokens)) * 100)
+      : null;
+
+  return {
+    health: health.status,
+    model,
+    contextUsage,
+    slots: slotsSummary,
+    metrics: {
+      generationSpeedTps: getMetricValue(metrics, "llamacpp:predicted_tokens_seconds"),
+      promptSpeedTps: getMetricValue(metrics, "llamacpp:prompt_tokens_seconds"),
+      avgGenerationSpeedTps: avgTokensPerSecond(
+        getMetricValue(metrics, "llamacpp:tokens_predicted_total"),
+        getMetricValue(metrics, "llamacpp:tokens_predicted_seconds_total"),
+      ),
+      avgPromptSpeedTps: avgTokensPerSecond(
+        getMetricValue(metrics, "llamacpp:prompt_tokens_total"),
+        getMetricValue(metrics, "llamacpp:prompt_seconds_total"),
+      ),
+      requestsProcessing: getMetricValue(metrics, "llamacpp:requests_processing"),
+      requestsDeferred: getMetricValue(metrics, "llamacpp:requests_deferred"),
+      tokensProcessed: getMetricValue(metrics, "llamacpp:prompt_tokens_total"),
+      tokensGenerated: getMetricValue(metrics, "llamacpp:tokens_predicted_total"),
+      promptCacheHitRate,
+      requestDecodedTokens: perRequest.decodedTokens,
+      taskId: perRequest.taskId,
+      specDraftTokens: getMetricValue(metrics, "llamacpp:spec_decode_num_draft_tokens_total"),
+      specAcceptedTokens: getMetricValue(metrics, "llamacpp:spec_decode_num_accepted_tokens_total"),
+      specDrafts: getMetricValue(metrics, "llamacpp:spec_decode_num_drafts_total"),
+    },
+  };
+};
+
+export const parseLlamacppHealthAsync = async (response: {
+  json: () => Promise<unknown>;
+}): Promise<z.infer<typeof llamacppHealthSchema>> => {
+  let json: unknown;
+  try {
+    json = await response.json();
+  } catch (error) {
+    throw new ParseError(LLMACPP_HEALTH_PARSE_ERROR_MESSAGE, {
+      cause: error instanceof Error ? error : new Error(String(error)),
+    });
+  }
+
+  const result = await llamacppHealthSchema.safeParseAsync(json);
+  if (!result.success) {
+    throw new ParseError(LLMACPP_HEALTH_PARSE_ERROR_MESSAGE, { cause: result.error });
+  }
+
+  return result.data;
+};
+
+export const parseLlamacppModelsAsync = async (response: {
+  json: () => Promise<unknown>;
+}): Promise<z.infer<typeof llamacppModelsSchema>> => {
+  let json: unknown;
+  try {
+    json = await response.json();
+  } catch (error) {
+    throw new ParseError(LLMACPP_MODELS_PARSE_ERROR_MESSAGE, {
+      cause: error instanceof Error ? error : new Error(String(error)),
+    });
+  }
+
+  const result = await llamacppModelsSchema.safeParseAsync(json);
+  if (!result.success) {
+    throw new ParseError(LLMACPP_MODELS_PARSE_ERROR_MESSAGE, { cause: result.error });
+  }
+
+  return result.data;
+};
+
+export const parseLlamacppMetricsAsync = async (response: {
+  text: () => Promise<string>;
+}): Promise<{ name: string; value: number }[]> => {
+  let text: string;
+  try {
+    text = await response.text();
+  } catch (error) {
+    throw new ParseError("Invalid llama.cpp /metrics response", {
+      cause: error instanceof Error ? error : new Error(String(error)),
+    });
+  }
+
+  return parsePrometheusMetrics(text);
+};

packages/integrations/src/llama-cpp/test/llamacpp-integration.spec.ts

@@ -0,0 +1,489 @@
+// @vitest-environment node
+import { ParseError } from "@homarr/common/server";
+import { Response } from "undici";
+import { beforeEach, describe, expect, test, vi } from "vitest";
+
+vi.hoisted(() => {
+  process.env.SKIP_ENV_VALIDATION = "true";
+  process.env.SECRET_ENCRYPTION_KEY = "ff3f4f7ce30e870c9630de9e5d244ffa81101a24ed0dfe5f064beb53a7e684f1";
+});
+
+import { fetchWithTrustedCertificatesAsync } from "@homarr/core/infrastructure/http";
+
+import type { IntegrationSecret } from "../../base/types";
+import { IntegrationParseError } from "../../base/errors/parse/integration-parse-error";
+import { LlamacppIntegration } from "../llamacpp-integration";
+import {
+  avgTokensPerSecond,
+  mapContextUsage,
+  mapLlamacppModel,
+  mapLlamacppPerRequest,
+  mapSlotsSummary,
+  parsePrometheusMetrics,
+  requestSpeedTps,
+} from "../llamacpp-types";
+
+vi.mock("@homarr/core/infrastructure/http", () => ({
+  fetchWithTrustedCertificatesAsync: vi.fn(),
+}));
+
+const TEST_URL = "http://llamacpp.example.com";
+const mockFetch = vi.mocked(fetchWithTrustedCertificatesAsync);
+
+const sampleModel = {
+  id: "/root/models/unsloth/Qwen3.8-27B-UD-Q4_K_XL.gguf",
+  meta: {
+    n_ctx: 131072,
+    n_params: 27320697856,
+    size: 17912397824,
+    ftype: "Q4_K - Small",
+  },
+};
+
+const sampleModelsResponse = {
+  data: [sampleModel],
+};
+
+const sampleMetricsText = [
+  "# HELP llamacpp:prompt_tokens_total Total number of tokens processed",
+  "# TYPE llamacpp:prompt_tokens_total counter",
+  "llamacpp:prompt_tokens_total 69834",
+  "llamacpp:prompt_seconds_total 23278",
+  "llamacpp:prompt_tokens_cached_total 12000",
+  "llamacpp:tokens_predicted_total 12161",
+  "llamacpp:tokens_predicted_seconds_total 380.03125",
+  "llamacpp:requests_processing 1",
+  "llamacpp:requests_deferred 0",
+  "llamacpp:predicted_tokens_seconds 32.8472",
+  "llamacpp:prompt_tokens_seconds 1500.5",
+  "llamacpp:spec_decode_num_draft_tokens_total 128",
+  "llamacpp:spec_decode_num_accepted_tokens_total 96",
+  "llamacpp:spec_decode_num_drafts_total 16",
+].join("\n");
+
+const sampleSlotsResponse = [
+  {
+    id: 0,
+    n_ctx: 131072,
+    n_prompt_tokens: 101076,
+    is_processing: true,
+    speculative: false,
+    id_task: 7,
+    next_token: [{ n_decoded: 42, n_remain: 100, has_next_token: true, has_new_line: false }],
+  },
+];
+
+const jsonResponse = (body: unknown, status = 200) =>
+  new Response(JSON.stringify(body), {
+    status,
+    headers: { "content-type": "application/json" },
+  }) as unknown as Awaited<ReturnType<typeof fetchWithTrustedCertificatesAsync>>;
+
+const textResponse = (body: string, status = 200) =>
+  new Response(body, {
+    status,
+    headers: { "content-type": "text/plain" },
+  }) as unknown as Awaited<ReturnType<typeof fetchWithTrustedCertificatesAsync>>;
+
+const createIntegration = (decryptedSecrets: IntegrationSecret[] = []) =>
+  new LlamacppIntegration({
+    id: "test-llamacpp",
+    name: "Test llama.cpp",
+    url: TEST_URL,
+    externalUrl: null,
+    decryptedSecrets,
+  });
+
+const mockAllEndpoints = () => {
+  mockFetch.mockReset();
+  mockFetch.mockImplementation((async (input) => {
+    const url = String(input);
+    if (url.includes("/health")) {
+      return jsonResponse({ status: "ok" });
+    }
+    if (url.includes("/v1/models")) {
+      return jsonResponse(sampleModelsResponse);
+    }
+    if (url.includes("/metrics")) {
+      return textResponse(sampleMetricsText);
+    }
+    if (url.includes("/slots")) {
+      return jsonResponse(sampleSlotsResponse);
+    }
+    return textResponse("not found", 404);
+  }) as typeof fetchWithTrustedCertificatesAsync);
+};
+
+beforeEach(() => {
+  mockFetch.mockReset();
+});
+
+describe("parsePrometheusMetrics", () => {
+  test("parses metric lines and skips comments and HELP/TYPE lines", () => {
+    const metrics = parsePrometheusMetrics(sampleMetricsText);
+
+    expect(metrics).toStrictEqual([
+      { name: "llamacpp:prompt_tokens_total", value: 69834 },
+      { name: "llamacpp:prompt_seconds_total", value: 23278 },
+      { name: "llamacpp:prompt_tokens_cached_total", value: 12000 },
+      { name: "llamacpp:tokens_predicted_total", value: 12161 },
+      { name: "llamacpp:tokens_predicted_seconds_total", value: 380.03125 },
+      { name: "llamacpp:requests_processing", value: 1 },
+      { name: "llamacpp:requests_deferred", value: 0 },
+      { name: "llamacpp:predicted_tokens_seconds", value: 32.8472 },
+      { name: "llamacpp:prompt_tokens_seconds", value: 1500.5 },
+      { name: "llamacpp:spec_decode_num_draft_tokens_total", value: 128 },
+      { name: "llamacpp:spec_decode_num_accepted_tokens_total", value: 96 },
+      { name: "llamacpp:spec_decode_num_drafts_total", value: 16 },
+    ]);
+  });
+
+  test("parses metric lines with labels and skips non-numeric values", () => {
+    const metrics = parsePrometheusMetrics(
+      [
+        'llamacpp:requests_by_status{status="ok"} 5',
+        "llamacpp:uptime_seconds +Inf",
+        "llamacpp:nan_metric NaN",
+        "",
+      ].join("\n"),
+    );
+
+    expect(metrics).toStrictEqual([{ name: "llamacpp:requests_by_status", value: 5 }]);
+  });
+
+  test("returns an empty array for empty input", () => {
+    expect(parsePrometheusMetrics("")).toStrictEqual([]);
+  });
+});
+
+describe("mapLlamacppModel", () => {
+  test("extracts a short display name and metadata fields", () => {
+    const model = mapLlamacppModel(sampleModel);
+
+    expect(model).toStrictEqual({
+      id: "/root/models/unsloth/Qwen3.8-27B-UD-Q4_K_XL.gguf",
+      name: "Qwen3.8-27B-UD-Q4_K_XL",
+      contextSize: 131072,
+      parameterCount: 27320697856,
+      fileSizeBytes: 17912397824,
+      quantization: "Q4_K - Small",
+    });
+  });
+
+  test("returns null metadata when meta is missing", () => {
+    const model = mapLlamacppModel({ id: "models/foo.gguf" });
+
+    expect(model).toStrictEqual({
+      id: "models/foo.gguf",
+      name: "foo",
+      contextSize: null,
+      parameterCount: null,
+      fileSizeBytes: null,
+      quantization: null,
+    });
+  });
+});
+
+describe("mapContextUsage", () => {
+  test("computes percent from a slot with valid n_ctx and n_prompt_tokens", () => {
+    expect(mapContextUsage(sampleSlotsResponse)).toStrictEqual({
+      usedTokens: 101076,
+      contextSize: 131072,
+      percent: 77.1,
+    });
+  });
+
+  test("returns the slot with the greatest used tokens, not the first valid slot", () => {
+    expect(
+      mapContextUsage([
+        { n_ctx: 1000, n_prompt_tokens: 100 },
+        { n_ctx: 8192, n_prompt_tokens: 6000 },
+        { n_ctx: 4096, n_prompt_tokens: 3000 },
+      ]),
+    ).toStrictEqual({
+      usedTokens: 6000,
+      contextSize: 8192,
+      percent: 73.2,
+    });
+  });
+
+  test("returns null for missing, non-array, or malformed slots", () => {
+    expect(mapContextUsage(null)).toBeNull();
+    expect(mapContextUsage([{ n_ctx: 0, n_prompt_tokens: 10 }])).toBeNull();
+    expect(mapContextUsage([{ n_prompt_tokens: 10 }])).toBeNull();
+    expect(mapContextUsage(["not-an-object", null])).toBeNull();
+  });
+
+  test("clamps the percentage to 100", () => {
+    expect(mapContextUsage([{ n_ctx: 10, n_prompt_tokens: 50 }])).toStrictEqual({
+      usedTokens: 50,
+      contextSize: 10,
+      percent: 100,
+    });
+  });
+});
+
+describe("avgTokensPerSecond", () => {
+  test("divides cumulative counters", () => {
+    expect(avgTokensPerSecond(12161, 380.03125)).toBe(32);
+  });
+
+  test("returns null when either counter is missing or seconds are zero", () => {
+    expect(avgTokensPerSecond(null, 5)).toBeNull();
+    expect(avgTokensPerSecond(10, null)).toBeNull();
+    expect(avgTokensPerSecond(10, 0)).toBeNull();
+  });
+});
+
+describe("mapLlamacppPerRequest", () => {
+  test("returns the active request id and decoded tokens from the processing slot", () => {
+    expect(mapLlamacppPerRequest(sampleSlotsResponse)).toStrictEqual({ taskId: 7, decodedTokens: 42 });
+  });
+
+  test("skips idle slots and returns nulls", () => {
+    expect(
+      mapLlamacppPerRequest([{ is_processing: false, id_task: 3, next_token: [{ n_decoded: 10 }] }]),
+    ).toStrictEqual({ taskId: null, decodedTokens: null });
+  });
+
+  test("returns nulls for missing, empty, or malformed slots", () => {
+    expect(mapLlamacppPerRequest(null)).toStrictEqual({ taskId: null, decodedTokens: null });
+    expect(mapLlamacppPerRequest([])).toStrictEqual({ taskId: null, decodedTokens: null });
+    expect(mapLlamacppPerRequest(["nope", null])).toStrictEqual({ taskId: null, decodedTokens: null });
+  });
+
+  test("handles a processing slot missing the per-request fields", () => {
+    expect(mapLlamacppPerRequest([{ is_processing: true }])).toStrictEqual({
+      taskId: null,
+      decodedTokens: null,
+    });
+  });
+});
+
+describe("mapSlotsSummary", () => {
+  test("counts slots, processing slots, decoding slots and speculative flags", () => {
+    expect(
+      mapSlotsSummary([
+        { is_processing: true, next_token: [{ n_decoded: 4 }], speculative: false },
+        { is_processing: false, speculative: true },
+        { is_processing: true, next_token: [], speculative: false },
+      ]),
+    ).toStrictEqual({ total: 3, processing: 2, decoding: 1, speculative: true });
+  });
+
+  test("treats an empty next_token list as not decoding", () => {
+    expect(mapSlotsSummary([{ is_processing: true, next_token: [] }, { is_processing: false }])).toStrictEqual({
+      total: 2,
+      processing: 1,
+      decoding: 0,
+      speculative: null,
+    });
+  });
+
+  test("returns null for missing, empty, or malformed slot lists", () => {
+    expect(mapSlotsSummary(null)).toBeNull();
+    expect(mapSlotsSummary([])).toBeNull();
+    expect(mapSlotsSummary(["not-an-object", null])).toBeNull();
+  });
+});
+
+describe("requestSpeedTps", () => {
+  test("divides token delta by elapsed seconds", () => {
+    expect(requestSpeedTps(128, 4)).toBe(32);
+  });
+
+  test("returns null for missing values, non-positive, or negative elapsed", () => {
+    expect(requestSpeedTps(null, 4)).toBeNull();
+    expect(requestSpeedTps(10, null)).toBeNull();
+    expect(requestSpeedTps(10, 0)).toBeNull();
+    expect(requestSpeedTps(10, -2)).toBeNull();
+  });
+});
+
+describe("LlamacppIntegration getStatsAsync", () => {
+  test("fetches health, models, metrics and slots and maps them into stats", async () => {
+    mockAllEndpoints();
+
+    const stats = await createIntegration().getStatsAsync();
+
+    expect(stats).toStrictEqual({
+      health: "ok",
+      model: {
+        id: "/root/models/unsloth/Qwen3.8-27B-UD-Q4_K_XL.gguf",
+        name: "Qwen3.8-27B-UD-Q4_K_XL",
+        contextSize: 131072,
+        parameterCount: 27320697856,
+        fileSizeBytes: 17912397824,
+        quantization: "Q4_K - Small",
+      },
+      contextUsage: {
+        usedTokens: 101076,
+        contextSize: 131072,
+        percent: 77.1,
+      },
+      slots: { total: 1, processing: 1, decoding: 1, speculative: false },
+      metrics: {
+        generationSpeedTps: 32.8472,
+        promptSpeedTps: 1500.5,
+        avgGenerationSpeedTps: 32,
+        avgPromptSpeedTps: 3,
+        requestsProcessing: 1,
+        requestsDeferred: 0,
+        tokensProcessed: 69834,
+        tokensGenerated: 12161,
+        promptCacheHitRate: 15,
+        requestDecodedTokens: 42,
+        taskId: 7,
+        specDraftTokens: 128,
+        specAcceptedTokens: 96,
+        specDrafts: 16,
+      },
+    });
+
+    const requestedUrls = mockFetch.mock.calls.map((call) => String(call[0])).toSorted();
+    expect(requestedUrls).toStrictEqual(
+      [`${TEST_URL}/health`, `${TEST_URL}/metrics`, `${TEST_URL}/slots`, `${TEST_URL}/v1/models`].toSorted(),
+    );
+  });
+
+  test("returns null model when the server has no models loaded", async () => {
+    mockAllEndpoints();
+    mockFetch.mockImplementation((async (input) => {
+      const url = String(input);
+      if (url.includes("/v1/models")) {
+        return jsonResponse({ data: [] });
+      }
+      if (url.includes("/health")) {
+        return jsonResponse({ status: "ok" });
+      }
+      if (url.includes("/slots")) {
+        return jsonResponse([]);
+      }
+      return textResponse(sampleMetricsText);
+    }) as typeof fetchWithTrustedCertificatesAsync);
+
+    const stats = await createIntegration().getStatsAsync();
+
+    expect(stats.model).toBeNull();
+    expect(stats.metrics.generationSpeedTps).toBe(32.8472);
+  });
+
+  test("returns null for metrics that are absent from /metrics", async () => {
+    mockAllEndpoints();
+    mockFetch.mockImplementation((async (input) => {
+      const url = String(input);
+      if (url.includes("/health")) {
+        return jsonResponse({ status: "ok" });
+      }
+      if (url.includes("/v1/models")) {
+        return jsonResponse({ data: [] });
+      }
+      if (url.includes("/slots")) {
+        return jsonResponse([]);
+      }
+      return textResponse("");
+    }) as typeof fetchWithTrustedCertificatesAsync);
+
+    const stats = await createIntegration().getStatsAsync();
+
+    expect(stats.metrics).toStrictEqual({
+      generationSpeedTps: null,
+      promptSpeedTps: null,
+      avgGenerationSpeedTps: null,
+      avgPromptSpeedTps: null,
+      requestsProcessing: null,
+      requestsDeferred: null,
+      tokensProcessed: null,
+      tokensGenerated: null,
+      promptCacheHitRate: null,
+      requestDecodedTokens: null,
+      taskId: null,
+      specDraftTokens: null,
+      specAcceptedTokens: null,
+      specDrafts: null,
+    });
+  });
+
+  test("throws when the /health endpoint returns an error status", async () => {
+    mockAllEndpoints();
+    mockFetch.mockImplementation((async (input) => {
+      const url = String(input);
+      if (url.includes("/health")) {
+        return textResponse("Service Unavailable", 503);
+      }
+      return jsonResponse({ status: "ok" });
+    }) as typeof fetchWithTrustedCertificatesAsync);
+
+    await expect(createIntegration().getStatsAsync()).rejects.toThrow();
+  });
+
+  test("degrades to null metrics when /metrics is not available (501) but keeps health and model data", async () => {
+    mockAllEndpoints();
+    mockFetch.mockImplementation((async (input) => {
+      const url = String(input);
+      if (url.includes("/metrics")) {
+        return textResponse("Not Implemented", 501);
+      }
+      if (url.includes("/health")) {
+        return jsonResponse({ status: "ok" });
+      }
+      if (url.includes("/v1/models")) {
+        return jsonResponse(sampleModelsResponse);
+      }
+      if (url.includes("/slots")) {
+        return jsonResponse(sampleSlotsResponse);
+      }
+      return textResponse("not found", 404);
+    }) as typeof fetchWithTrustedCertificatesAsync);
+
+    const stats = await createIntegration().getStatsAsync();
+
+    expect(stats.health).toBe("ok");
+    expect(stats.model?.name).toBe("Qwen3.8-27B-UD-Q4_K_XL");
+    expect(stats.contextUsage).toStrictEqual({
+      usedTokens: 101076,
+      contextSize: 131072,
+      percent: 77.1,
+    });
+    expect(stats.slots).toStrictEqual({ total: 1, processing: 1, decoding: 1, speculative: false });
+    expect(stats.metrics).toStrictEqual({
+      generationSpeedTps: null,
+      promptSpeedTps: null,
+      avgGenerationSpeedTps: null,
+      avgPromptSpeedTps: null,
+      requestsProcessing: null,
+      requestsDeferred: null,
+      tokensProcessed: null,
+      tokensGenerated: null,
+      promptCacheHitRate: null,
+      requestDecodedTokens: 42,
+      taskId: 7,
+      specDraftTokens: null,
+      specAcceptedTokens: null,
+      specDrafts: null,
+    });
+  });
+
+  test("throws a parse error when /v1/models is not valid JSON", async () => {
+    mockAllEndpoints();
+    mockFetch.mockImplementation((async (input) => {
+      const url = String(input);
+      if (url.includes("/v1/models")) {
+        return textResponse("not-json");
+      }
+      if (url.includes("/health")) {
+        return jsonResponse({ status: "ok" });
+      }
+      if (url.includes("/slots")) {
+        return jsonResponse([]);
+      }
+      return textResponse(sampleMetricsText);
+    }) as typeof fetchWithTrustedCertificatesAsync);
+
+    await expect(createIntegration().getStatsAsync()).rejects.toSatisfy((error) => {
+      if (!(error instanceof IntegrationParseError)) return false;
+      const cause = error.cause;
+      return cause instanceof ParseError && cause.message.includes("Invalid llama.cpp /v1/models response");
+    });
+  });
+});

packages/integrations/src/mock/widget-data.ts

@@ -11,6 +11,7 @@ import type { TracearrDashboardData } from "../tracearr/tracearr-types";
 import type { TraefikDashboardData } from "../traefik/traefik-types";
 import type { UmamiEventSeries, UmamiMetricItem, UmamiVisitorStats, UmamiWebsite } from "../umami/umami-types";
 import type { UptimeKumaDashboardData } from "../uptime-kuma/uptime-kuma-types";
+import type { LlamacppStats } from "../llama-cpp/llamacpp-types";
 import type { WudStats } from "../wud/wud-types";
 import type {
   FirewallCpuSummary,
@@ -352,4 +353,33 @@ export const mockWidgetData = {
       { id: "redis", name: "Redis", currentVersion: "8.0.2", newVersion: "8.0.3", link: null },
     ],
   } satisfies WudStats,
+  llamacpp: {
+    health: "ok",
+    model: {
+      id: "unsloth/Qwen3.8-27B-UD-Q4_K_XL.gguf",
+      name: "Qwen3.8-27B-UD-Q4_K_XL",
+      contextSize: 32768,
+      parameterCount: 27_347_098_624,
+      fileSizeBytes: 16_842_776_576,
+      quantization: "Q4_K_XL",
+    },
+    contextUsage: { usedTokens: 8192, contextSize: 32768, percent: 25 },
+    slots: { total: 2, processing: 1, decoding: 1, speculative: false },
+    metrics: {
+      generationSpeedTps: 42.5,
+      promptSpeedTps: 1840.2,
+      avgGenerationSpeedTps: 39.8,
+      avgPromptSpeedTps: 2104.6,
+      requestsProcessing: 1,
+      requestsDeferred: 0,
+      tokensProcessed: 12_483,
+      tokensGenerated: 1_024,
+      promptCacheHitRate: 62,
+      requestDecodedTokens: 96,
+      taskId: 1042,
+      specDraftTokens: 48,
+      specAcceptedTokens: 12,
+      specDrafts: 8,
+    },
+  } satisfies LlamacppStats,
 } as const;

packages/request-handler/src/llamacpp.ts

@@ -0,0 +1,16 @@
+import { createIntegrationAsync } from "@homarr/integrations/factory";
+import type { LlamacppStats } from "@homarr/integrations";
+
+import { createIntegrationRequestHandler } from "./lib/integration-request-handler";
+
+export const llamacppStatsRequestHandler = createIntegrationRequestHandler<
+  LlamacppStats,
+  "llamacpp",
+  Record<string, never>
+>({
+  cacheNamespace: "llamacpp:stats",
+  async requestAsync(integration) {
+    const integrationInstance = await createIntegrationAsync(integration);
+    return await integrationInstance.getStatsAsync();
+  },
+});

packages/translation/src/lang/cn.json

@@ -1027,7 +1027,8 @@
       "analytics": "分析",
       "vpn": "VPN",
       "mediaLibrary": "媒体库",
-      "uptimeMonitoring": "运行时间监控"
+      "uptimeMonitoring": "运行时间监控",
+      "systemMonitoring": "系统监控"
     },
     "testConnection": {
       "action": {
@@ -1355,6 +1356,10 @@
     "bazarr": {
       "name": "Bazarr",
       "description": "Sonarr 和 Radarr 的配套应用，用于管理和下载字幕"
+    },
+    "llamacpp": {
+      "name": "llama.cpp",
+      "description": "监控本地 llama.cpp llama-server：健康状态、模型与生成指标"
     }
   },
   "media": {
@@ -4171,6 +4176,81 @@
           "subtitle": "Docker 容器的网络流量"
         }
       }
+    },
+    "llamacpp": {
+      "name": "llama.cpp",
+      "description": "显示本地 llama.cpp 服务器的健康状态、加载模型和生成速度。",
+      "title": "llama.cpp",
+      "status": {
+        "unhealthy": "异常 ({status})"
+      },
+      "idle": "空闲",
+      "speedTooltip": {
+        "current": "当前生成速度",
+        "average": "启动以来的平均生成速度",
+        "request": "本次请求开始以来的平均速度"
+      },
+      "busy": "{count, plural, other {# 个请求处理中}}",
+      "modelInfo": {
+        "context": "上下文",
+        "size": "文件大小",
+        "parameters": "参数量",
+        "model": "模型",
+        "quantization": "量化"
+      },
+      "stats": {
+        "tokensProcessed": "已处理 Token",
+        "contextUsage": "上下文 / KV 占用",
+        "cacheHitRate": "提示缓存命中率",
+        "tokensGenerated": "已生成 Token"
+      },
+      "error": {
+        "unreachable": "无法连接 llama-server"
+      },
+      "unit": {
+        "speed": "{value} t/s",
+        "parameters": "{value}B",
+        "none": "—",
+        "speedShort": "t/s"
+      },
+      "option": {
+        "showTitle": {
+          "label": "显示标题"
+        },
+        "showModelInfo": {
+          "label": "显示模型信息"
+        },
+        "showContextUsage": {
+          "label": "显示上下文占用"
+        },
+        "showCacheHitRate": {
+          "label": "显示缓存命中率"
+        },
+        "showSpeedStats": {
+          "label": "显示速度统计"
+        },
+        "showRequests": {
+          "label": "显示请求负载"
+        },
+        "showTokenStats": {
+          "label": "显示 Token 总量"
+        },
+        "showSpeculative": {
+          "label": "显示投机解码"
+        }
+      },
+      "advanced": {
+        "promptSpeed": "Prompt 速度",
+        "avgSpeed": "平均生成速度",
+        "requests": "请求数",
+        "deferred": "排队中",
+        "slots": "Slot 占用 {processing}/{total}",
+        "speculative": "投机解码",
+        "avgPrompt": "平均 Prompt 速度",
+        "specDrafts": "投机草稿数",
+        "specAccepted": "投机接受 Token",
+        "specDraftTokens": "投机草稿 Token"
+      }
     }
   },
   "widgetPreview": {

packages/translation/src/lang/en.json

@@ -1424,7 +1424,8 @@
       "vpn": "VPN",
       "mediaLibrary": "Media Library",
       "uptimeMonitoring": "Uptime Monitoring",
-      "beszel": "Beszel"
+      "beszel": "Beszel",
+      "systemMonitoring": "System Monitoring"
     },
     "testConnection": {
       "action": {
@@ -1716,6 +1717,10 @@
       "use": "Select integrations in items",
       "interact": "Interact with integrations",
       "full": "Full integration access"
+    },
+    "llamacpp": {
+      "name": "llama.cpp",
+      "description": "Monitor a local llama.cpp llama-server: health, model and generation metrics"
     }
   },
   "media": {
@@ -5205,6 +5210,81 @@
         "uptime": "Uptime",
         "agent": "Agent"
       }
+    },
+    "llamacpp": {
+      "name": "llama.cpp",
+      "description": "Show the health, loaded model and generation speed of your local llama.cpp server.",
+      "title": "llama.cpp",
+      "status": {
+        "unhealthy": "Unhealthy ({status})"
+      },
+      "idle": "Idle",
+      "speedTooltip": {
+        "current": "Current generation speed",
+        "average": "Average generation speed since startup",
+        "request": "Average speed since this request started"
+      },
+      "busy": "{count, plural, one {# request running} other {# requests running}}",
+      "modelInfo": {
+        "context": "Context",
+        "size": "File size",
+        "parameters": "Parameters",
+        "model": "Model",
+        "quantization": "Quantization"
+      },
+      "stats": {
+        "tokensProcessed": "Tokens processed",
+        "contextUsage": "Context / KV usage",
+        "cacheHitRate": "Prompt cache hit rate",
+        "tokensGenerated": "Tokens generated"
+      },
+      "error": {
+        "unreachable": "llama-server is not reachable"
+      },
+      "unit": {
+        "speed": "{value} t/s",
+        "parameters": "{value}B",
+        "none": "—",
+        "speedShort": "t/s"
+      },
+      "option": {
+        "showTitle": {
+          "label": "Show title"
+        },
+        "showModelInfo": {
+          "label": "Show model info"
+        },
+        "showContextUsage": {
+          "label": "Show context usage"
+        },
+        "showCacheHitRate": {
+          "label": "Show cache hit rate"
+        },
+        "showSpeedStats": {
+          "label": "Show speed stats"
+        },
+        "showRequests": {
+          "label": "Show request load"
+        },
+        "showTokenStats": {
+          "label": "Show token totals"
+        },
+        "showSpeculative": {
+          "label": "Show speculative decoding"
+        }
+      },
+      "advanced": {
+        "promptSpeed": "Prompt speed",
+        "avgSpeed": "Avg generation",
+        "requests": "Requests",
+        "deferred": "Deferred",
+        "slots": "{processing}/{total} slots busy",
+        "speculative": "Speculative decoding",
+        "avgPrompt": "Avg prompt",
+        "specDrafts": "Spec. drafts",
+        "specAccepted": "Spec. accepted tokens",
+        "specDraftTokens": "Spec. draft tokens"
+      }
     }
   },
   "widgetPreview": {
@@ -6157,6 +6237,9 @@
                 "lm-studio": {
                   "description": "An LM Studio local server with OpenAI compatibility enabled."
                 },
+                "llama-cpp": {
+                  "description": "A local or network llama.cpp server using its OpenAI-compatible API."
+                },
                 "custom": {
                   "label": "Custom endpoint",
                   "description": "Any other endpoint compatible with OpenAI Chat Completions."

packages/translation/src/lang/zh.json

@@ -822,7 +822,8 @@
       "analytics": "分析",
       "vpn": "VPN",
       "mediaLibrary": "媒體庫",
-      "uptimeMonitoring": "運作時間監控"
+      "uptimeMonitoring": "運作時間監控",
+      "systemMonitoring": "系統監控"
     },
     "testConnection": {
       "action": {
@@ -1150,6 +1151,10 @@
     "bazarr": {
       "name": "Bazarr",
       "description": "Sonarr 和 Radarr 的配套應用程式，管理和下載字幕"
+    },
+    "llamacpp": {
+      "name": "llama.cpp",
+      "description": "監控本地 llama.cpp llama-server：健康狀態、模型與生成指標"
     }
   },
   "media": {
@@ -3946,6 +3951,81 @@
       "error": {
         "internalServerError": "無法擷取 Traefik 資料"
       }
+    },
+    "llamacpp": {
+      "name": "llama.cpp",
+      "description": "顯示本地 llama.cpp 伺服器的健康狀態、載入模型與生成速度。",
+      "title": "llama.cpp",
+      "status": {
+        "unhealthy": "異常 ({status})"
+      },
+      "idle": "閒置",
+      "speedTooltip": {
+        "current": "目前生成速度",
+        "average": "啟動以來的平均生成速度",
+        "request": "本次請求開始以來的平均速度"
+      },
+      "busy": "{count, plural, other {# 個請求處理中}}",
+      "modelInfo": {
+        "context": "上下文",
+        "size": "檔案大小",
+        "parameters": "參數量",
+        "model": "模型",
+        "quantization": "量化"
+      },
+      "stats": {
+        "tokensProcessed": "已處理 Token",
+        "contextUsage": "上下文 / KV 佔用",
+        "cacheHitRate": "提示快取命中率",
+        "tokensGenerated": "已生成 Token"
+      },
+      "error": {
+        "unreachable": "無法連接 llama-server"
+      },
+      "unit": {
+        "speed": "{value} t/s",
+        "parameters": "{value}B",
+        "none": "—",
+        "speedShort": "t/s"
+      },
+      "option": {
+        "showTitle": {
+          "label": "顯示標題"
+        },
+        "showModelInfo": {
+          "label": "顯示模型資訊"
+        },
+        "showContextUsage": {
+          "label": "顯示上下文佔用"
+        },
+        "showCacheHitRate": {
+          "label": "顯示快取命中率"
+        },
+        "showSpeedStats": {
+          "label": "顯示速度統計"
+        },
+        "showRequests": {
+          "label": "顯示請求負載"
+        },
+        "showTokenStats": {
+          "label": "顯示 Token 總量"
+        },
+        "showSpeculative": {
+          "label": "顯示投機解碼"
+        }
+      },
+      "advanced": {
+        "promptSpeed": "Prompt 速度",
+        "avgSpeed": "平均生成速度",
+        "requests": "請求數",
+        "deferred": "排隊中",
+        "slots": "Slot 佔用 {processing}/{total}",
+        "speculative": "投機解碼",
+        "avgPrompt": "平均 Prompt 速度",
+        "specDrafts": "投機草稿數",
+        "specAccepted": "投機接受 Token",
+        "specDraftTokens": "投機草稿 Token"
+      }
     }
   },
   "widgetPreview": {

packages/ui/src/widget-icons.ts

@@ -20,6 +20,7 @@ import {
   IconClock,
   IconClockPlay,
   IconCloud,
+  IconCpu,
   IconDeviceCctv,
   IconDeviceGamepad,
   IconDownload,
@@ -116,4 +117,5 @@ export const widgetCatalogIcons: Record<WidgetKind, TablerIcon> = {
   traefik: IconRoute,
   customApi: IconApi,
   wud: IconBrandDocker,
+  llamacpp: IconCpu,
 };

packages/widgets/src/llama-cpp/component.tsx

@@ -0,0 +1,571 @@
+"use client";
+
+import { useMemo, useRef } from "react";
+import type { ReactNode } from "react";
+
+import { Box, Center, Group, Progress, RingProgress, ScrollArea, Stack, Text, Tooltip } from "@mantine/core";
+import {
+  IconBrain,
+  IconBolt,
+  IconCircleX,
+  IconDatabase,
+  IconFileText,
+  IconGauge,
+  IconLayersIntersect,
+  IconServer,
+} from "@tabler/icons-react";
+
+import { clientApi } from "@homarr/api/client";
+import { formatBytes, formatNumber } from "@homarr/common";
+import { useI18n } from "@homarr/translation/client";
+
+import { WidgetEmptyState } from "../common/empty-state";
+import type { WidgetComponentProps } from "../definition";
+import { NoIntegrationSelectedError } from "../errors/no-integration-selected";
+
+// Board grid: a widget N cells wide is N*200 + (N-1)*12 px (1→200, 2→412, 3→624, 4→836).
+// Width drives how much info fits side-by-side; height only matters for the
+// single-row (short) case where the widget must drop down to essentials.
+const isNarrowWidth = (width: number) => width < 300; // 1 cell (200px)
+const isWideWidth = (width: number) => width >= 550; // 3+ cells (624px+)
+const isShortHeight = (height: number) => height < 300; // 1 row (200px)
+
+const ringSizeByScale = [
+  { minWidth: 320, size: 108 },
+  { minWidth: 220, size: 92 },
+  { minWidth: 0, size: 68 },
+] as const;
+
+const getContextColor = (percent: number) => (percent > 90 ? "red" : percent > 75 ? "yellow" : "blue");
+
+export default function LlamacppWidget({
+  integrationIds,
+  options,
+  width,
+  height,
+  displayMode = "compact",
+}: WidgetComponentProps<"llamacpp">) {
+  const integrationId = integrationIds[0];
+  if (!integrationId) {
+    throw new NoIntegrationSelectedError();
+  }
+
+  return (
+    <LlamacppContent
+      integrationId={integrationId}
+      options={options}
+      width={width}
+      height={height}
+      displayMode={displayMode}
+    />
+  );
+}
+
+interface LlamacppContentProps {
+  integrationId: string;
+  options: WidgetComponentProps<"llamacpp">["options"];
+  width: number;
+  height: number;
+  displayMode: "compact" | "advanced";
+}
+
+function LlamacppContent({ integrationId, options, width, height, displayMode }: LlamacppContentProps) {
+  const t = useI18n("widget.llamacpp");
+  const { data, isError } = clientApi.widget.llamacpp.getStats.useQuery({ integrationId });
+
+  // Tracks the in-flight request across polls so the widget can show the average
+  // speed since that request started: /slots reports the request id (taskId) and how
+  // many tokens it has generated so far (requestDecodedTokens). The server resets that
+  // counter to 0 when a new request starts, so a fresh taskId means a fresh request.
+  const requestRef = useRef<{ taskId: number; baseTokens: number; baseTimeMs: number } | null>(null);
+  const perRequestSpeedTps = useMemo(() => {
+    const taskId = data?.stats.metrics.taskId ?? null;
+    const decoded = data?.stats.metrics.requestDecodedTokens ?? null;
+    if (taskId === null || decoded === null) {
+      requestRef.current = null;
+      return null;
+    }
+
+    const nowMs = Date.now();
+    const current = requestRef.current;
+    if (current === null || current.taskId !== taskId) {
+      requestRef.current = { taskId, baseTokens: decoded, baseTimeMs: nowMs };
+      return null;
+    }
+
+    const elapsedSeconds = (nowMs - current.baseTimeMs) / 1000;
+    const deltaTokens = decoded - current.baseTokens;
+    if (elapsedSeconds <= 0 || deltaTokens < 0) {
+      return null;
+    }
+    const speed = deltaTokens / elapsedSeconds;
+    return Number.isFinite(speed) && speed > 0 ? speed : null;
+  }, [data]);
+
+  if (isError) {
+    return (
+      <Center h="100%" w="100%" p="sm">
+        <Group gap="xs" wrap="nowrap" miw={0}>
+          <IconCircleX size={20} color="red" />
+          <Text size="sm" c="dimmed" lineClamp={1}>
+            {t("error.unreachable")}
+          </Text>
+        </Group>
+      </Center>
+    );
+  }
+
+  if (!data) {
+    return <WidgetEmptyState />;
+  }
+
+  const stats = data.stats;
+  const isHealthy = stats.health === "ok";
+  const model = stats.model;
+  const contextUsage = stats.contextUsage;
+  const slots = stats.slots;
+
+  const requestsProcessing = stats.metrics.requestsProcessing;
+  const isBusy = requestsProcessing !== null && requestsProcessing > 0;
+
+  let displayedSpeed: number | null;
+  let speedMode: "request" | "live" | "average";
+  if (isBusy && perRequestSpeedTps !== null) {
+    displayedSpeed = perRequestSpeedTps;
+    speedMode = "request";
+  } else if (isBusy && (stats.metrics.generationSpeedTps ?? 0) > 0) {
+    displayedSpeed = stats.metrics.generationSpeedTps;
+    speedMode = "live";
+  } else {
+    displayedSpeed = stats.metrics.avgGenerationSpeedTps;
+    speedMode = "average";
+  }
+  const speedTooltip =
+    speedMode === "request"
+      ? t("speedTooltip.request")
+      : speedMode === "live"
+        ? t("speedTooltip.current")
+        : t("speedTooltip.average");
+
+  const speedValue = displayedSpeed !== null && displayedSpeed > 0 ? formatNumber(displayedSpeed, 1) : t("unit.none");
+  const statusColor = isHealthy ? "green" : "red";
+  const statusLabel = isHealthy ? undefined : t("status.unhealthy", { status: stats.health });
+
+  const contextPercent = options.showContextUsage && contextUsage ? contextUsage.percent : null;
+  const contextLabel = contextUsage
+    ? `${formatNumber(contextUsage.usedTokens, 0)} / ${formatNumber(contextUsage.contextSize, 0)} (${contextUsage.percent}%)`
+    : null;
+
+  const ringSize = ringSizeByScale.find(({ minWidth }) => Math.min(width, height) >= minWidth)?.size ?? 68;
+  const ring = (
+    <Tooltip label={speedTooltip} withArrow position="bottom">
+      <RingProgress
+        size={ringSize}
+        thickness={Math.max(6, Math.round(ringSize / 10))}
+        roundCaps
+        sections={[
+          { value: contextPercent ?? 0, color: contextPercent !== null ? getContextColor(contextPercent) : "gray" },
+        ]}
+        label={
+          <Center>
+            <Stack gap={0} align="center" style={{ lineHeight: 1.15 }}>
+              <Text size={ringSize >= 92 ? "lg" : "sm"} fw={800} lineClamp={1}>
+                {speedValue}
+              </Text>
+              <Text size="xs" c="dimmed" lineClamp={1}>
+                {t("unit.speedShort")}
+              </Text>
+            </Stack>
+          </Center>
+        }
+      />
+    </Tooltip>
+  );
+
+  const modelLine = model && (
+    <Group gap={6} wrap="nowrap" miw={0}>
+      <IconBrain size={14} color="dimmed" style={{ flexShrink: 0 }} />
+      <Tooltip label={model.id} withArrow>
+        <Text size="xs" lineClamp={1} style={{ maxWidth: "100%" }}>
+          {model.name}
+        </Text>
+      </Tooltip>
+      {model.quantization && (
+        <Text size="xs" c="dimmed" fw={600} lineClamp={1} style={{ flexShrink: 0 }}>
+          {model.quantization}
+        </Text>
+      )}
+    </Group>
+  );
+
+  const busyLabel =
+    requestsProcessing !== null ? (isBusy ? t("busy", { count: requestsProcessing }) : t("idle")) : null;
+
+  const promptSpeed = stats.metrics.promptSpeedTps;
+  const avgPromptSpeed = stats.metrics.avgPromptSpeedTps;
+  const cacheHitRate = stats.metrics.promptCacheHitRate;
+  const tokensProcessed = stats.metrics.tokensProcessed;
+  const tokensGenerated = stats.metrics.tokensGenerated;
+  const requestsDeferred = stats.metrics.requestsDeferred;
+  const specDraftTokens = stats.metrics.specDraftTokens;
+  const specAcceptedTokens = stats.metrics.specAcceptedTokens;
+  const specDrafts = stats.metrics.specDrafts;
+  const hasSpecStats = specDraftTokens !== null || specAcceptedTokens !== null || specDrafts !== null;
+
+  if (displayMode === "advanced") {
+    return (
+      <Box h="100%" w="100%">
+        <ScrollArea h="100%">
+          <Stack gap="md" p="md">
+            <Group gap="xs" wrap="nowrap" miw={0}>
+              <StatusDot color={statusColor} label={statusLabel} />
+              <Text size="md" fw={700} lineClamp={1}>
+                {t("title")}
+              </Text>
+              {options.showRequests && busyLabel && (
+                <Text size="xs" c="dimmed" lineClamp={1}>
+                  {busyLabel}
+                </Text>
+              )}
+            </Group>
+
+            <Group gap="lg" wrap="nowrap" miw={0}>
+              {ring}
+              <Stack gap="sm" style={{ flex: 1, minWidth: 0 }}>
+                {options.showSpeedStats && (
+                  <AdvancedRow
+                    icon={<IconBolt size={14} color="dimmed" />}
+                    label={t("advanced.promptSpeed")}
+                    value={
+                      promptSpeed !== null && promptSpeed > 0
+                        ? t("unit.speed", { value: formatNumber(promptSpeed, 1) })
+                        : t("unit.none")
+                    }
+                  />
+                )}
+                {options.showSpeedStats && (
+                  <AdvancedRow
+                    icon={<IconBolt size={14} color="dimmed" />}
+                    label={t("advanced.avgPrompt")}
+                    value={
+                      avgPromptSpeed !== null && avgPromptSpeed > 0
+                        ? t("unit.speed", { value: formatNumber(avgPromptSpeed, 1) })
+                        : t("unit.none")
+                    }
+                  />
+                )}
+                {options.showSpeedStats && stats.metrics.avgGenerationSpeedTps !== null && (
+                  <AdvancedRow
+                    icon={<IconGauge size={14} color="dimmed" />}
+                    label={t("advanced.avgSpeed")}
+                    value={
+                      (stats.metrics.avgGenerationSpeedTps ?? 0) > 0
+                        ? t("unit.speed", { value: formatNumber(stats.metrics.avgGenerationSpeedTps, 1) })
+                        : t("unit.none")
+                    }
+                  />
+                )}
+                {options.showContextUsage && contextUsage && (
+                  <AdvancedRow
+                    icon={<IconDatabase size={14} color="dimmed" />}
+                    label={t("stats.contextUsage")}
+                    value={contextLabel ?? "—"}
+                  />
+                )}
+                {options.showRequests && requestsProcessing !== null && (
+                  <AdvancedRow
+                    icon={<IconServer size={14} color="dimmed" />}
+                    label={t("advanced.requests")}
+                    value={formatNumber(requestsProcessing, 0)}
+                  />
+                )}
+                {options.showRequests && requestsDeferred !== null && (
+                  <AdvancedRow
+                    icon={<IconLayersIntersect size={14} color="dimmed" />}
+                    label={t("advanced.deferred")}
+                    value={formatNumber(requestsDeferred, 0)}
+                  />
+                )}
+              </Stack>
+            </Group>
+
+            {(options.showRequests || (options.showSpeculative && slots?.speculative)) && slots && (
+              <Group gap="sm" wrap="nowrap" miw={0}>
+                {options.showRequests && (
+                  <Text size="xs" c="dimmed" lineClamp={1}>
+                    {t("advanced.slots", { processing: slots.processing, total: slots.total })}
+                  </Text>
+                )}
+                {options.showSpeculative && slots.speculative && (
+                  <Text size="xs" c="dimmed" lineClamp={1} fw={600}>
+                    {t("advanced.speculative")}
+                  </Text>
+                )}
+              </Group>
+            )}
+
+            {options.showModelInfo && model && (
+              <Stack gap="xs">
+                <AdvancedRow
+                  icon={<IconBrain size={14} color="dimmed" />}
+                  label={t("modelInfo.model")}
+                  value={model.name}
+                  valueTooltip={model.id}
+                />
+                {model.quantization && (
+                  <AdvancedRow
+                    icon={<IconFileText size={14} color="dimmed" />}
+                    label={t("modelInfo.quantization")}
+                    value={model.quantization}
+                  />
+                )}
+                {model.contextSize !== null && (
+                  <AdvancedRow
+                    icon={<IconDatabase size={14} color="dimmed" />}
+                    label={t("modelInfo.context")}
+                    value={formatNumber(model.contextSize, 0)}
+                  />
+                )}
+                {model.fileSizeBytes !== null && (
+                  <AdvancedRow
+                    icon={<IconFileText size={14} color="dimmed" />}
+                    label={t("modelInfo.size")}
+                    value={formatBytes(model.fileSizeBytes)}
+                  />
+                )}
+                {model.parameterCount !== null && (
+                  <AdvancedRow
+                    icon={<IconBrain size={14} color="dimmed" />}
+                    label={t("modelInfo.parameters")}
+                    value={t("unit.parameters", { value: formatNumber(model.parameterCount / 1e9, 2) })}
+                  />
+                )}
+              </Stack>
+            )}
+
+            {options.showTokenStats && (tokensProcessed !== null || tokensGenerated !== null) && (
+              <Group gap="md" wrap="nowrap" miw={0}>
+                {tokensProcessed !== null && (
+                  <AdvancedRow
+                    icon={<IconFileText size={14} color="dimmed" />}
+                    label={t("stats.tokensProcessed")}
+                    value={formatNumber(tokensProcessed, 0)}
+                  />
+                )}
+                {tokensGenerated !== null && (
+                  <AdvancedRow
+                    icon={<IconBolt size={14} color="dimmed" />}
+                    label={t("stats.tokensGenerated")}
+                    value={formatNumber(tokensGenerated, 0)}
+                  />
+                )}
+              </Group>
+            )}
+
+            {options.showSpeculative && hasSpecStats && (
+              <Stack gap="xs">
+                {specDrafts !== null && (
+                  <AdvancedRow
+                    icon={<IconLayersIntersect size={14} color="dimmed" />}
+                    label={t("advanced.specDrafts")}
+                    value={formatNumber(specDrafts, 0)}
+                  />
+                )}
+                {specDraftTokens !== null && (
+                  <AdvancedRow
+                    icon={<IconFileText size={14} color="dimmed" />}
+                    label={t("advanced.specDraftTokens")}
+                    value={formatNumber(specDraftTokens, 0)}
+                  />
+                )}
+                {specAcceptedTokens !== null && (
+                  <AdvancedRow
+                    icon={<IconBolt size={14} color="dimmed" />}
+                    label={t("advanced.specAccepted")}
+                    value={formatNumber(specAcceptedTokens, 0)}
+                  />
+                )}
+              </Stack>
+            )}
+
+            {options.showCacheHitRate && cacheHitRate !== null && (
+              <AdvancedRow
+                icon={<IconDatabase size={14} color="dimmed" />}
+                label={t("stats.cacheHitRate")}
+                value={`${cacheHitRate}%`}
+                progress={{ value: cacheHitRate, color: cacheHitRate > 75 ? "green" : "blue" }}
+              />
+            )}
+          </Stack>
+        </ScrollArea>
+      </Box>
+    );
+  }
+
+  // Compact (default) mode — density adapts to the widget's grid footprint.
+  if (isNarrowWidth(width)) {
+    return (
+      <Center h="100%" w="100%">
+        <Stack gap={isShortHeight(height) ? 0 : 4} align="center" p="xs" style={{ position: "relative" }}>
+          <Center pos="relative">
+            {ring}
+            <Box pos="absolute" top={0} right={0} style={{ zIndex: 1 }}>
+              <StatusDot color={statusColor} label={statusLabel} />
+            </Box>
+          </Center>
+          {!isShortHeight(height) && options.showModelInfo && modelLine}
+        </Stack>
+      </Center>
+    );
+  }
+
+  const rightColumn = (
+    <Stack gap="xs" style={{ flex: 1, minWidth: 0 }}>
+      {options.showModelInfo && modelLine}
+      {options.showRequests && busyLabel && (
+        <Text size="xs" c="dimmed" lineClamp={1}>
+          {busyLabel}
+        </Text>
+      )}
+      {options.showContextUsage && contextUsage && (
+        <Group gap={6} wrap="nowrap" miw={0} w="100%" align="center">
+          <Progress
+            value={contextUsage.percent}
+            size="xs"
+            radius="xs"
+            color={getContextColor(contextUsage.percent)}
+            style={{ flex: 1, minWidth: 0 }}
+          />
+          <Text size="xs" c="dimmed" lineClamp={1} style={{ flexShrink: 0 }}>
+            {contextUsage.percent}%
+          </Text>
+        </Group>
+      )}
+      {isWideWidth(width) && (
+        <Stack gap="xs">
+          {options.showSpeedStats && promptSpeed !== null && promptSpeed > 0 && (
+            <InfoRow
+              label={t("advanced.promptSpeed")}
+              value={t("unit.speed", { value: formatNumber(promptSpeed, 1) })}
+            />
+          )}
+          {options.showCacheHitRate && cacheHitRate !== null && (
+            <InfoRow label={t("stats.cacheHitRate")} value={`${cacheHitRate}%`} />
+          )}
+          {options.showTokenStats && (tokensProcessed !== null || tokensGenerated !== null) && (
+            <InfoRow
+              label={t("stats.tokensProcessed")}
+              value={
+                tokensProcessed !== null && tokensGenerated !== null
+                  ? formatNumber(tokensProcessed + tokensGenerated, 0)
+                  : tokensProcessed !== null
+                    ? formatNumber(tokensProcessed, 0)
+                    : "—"
+              }
+            />
+          )}
+        </Stack>
+      )}
+    </Stack>
+  );
+
+  if (isShortHeight(height)) {
+    return (
+      <Center h="100%" w="100%">
+        <Group gap="md" w="100%" wrap="nowrap" miw={0} justify="center" p="xs">
+          {ring}
+          {!isNarrowWidth(width) && (
+            <Stack gap={2} style={{ minWidth: 0 }}>
+              {options.showModelInfo && modelLine}
+              {options.showRequests && busyLabel && (
+                <Text size="xs" c="dimmed" lineClamp={1}>
+                  {busyLabel}
+                </Text>
+              )}
+            </Stack>
+          )}
+        </Group>
+      </Center>
+    );
+  }
+
+  return (
+    <Box h="100%" p="xs" pos="relative">
+      <Group pos="absolute" top={6} right={10} gap={0} style={{ zIndex: 2 }}>
+        <StatusDot color={statusColor} label={statusLabel} />
+      </Group>
+      <Center h="100%">
+        <Group gap="md" w="100%" wrap="nowrap" miw={0}>
+          {ring}
+          {rightColumn}
+        </Group>
+      </Center>
+    </Box>
+  );
+}
+
+const StatusDot = ({ color, label }: { color: string; label?: string }) => {
+  const dot = (
+    <Box
+      aria-label={label}
+      style={{
+        width: 9,
+        height: 9,
+        borderRadius: "50%",
+        backgroundColor: `var(--mantine-color-${color}-6)`,
+        flexShrink: 0,
+        cursor: "default",
+      }}
+    />
+  );
+  if (!label) return dot;
+  return (
+    <Tooltip label={label} withArrow position="left">
+      {dot}
+    </Tooltip>
+  );
+};
+
+interface AdvancedRowProps {
+  icon: ReactNode;
+  label: string;
+  value: string;
+  valueTooltip?: string;
+  progress?: { value: number; color: string };
+}
+
+const AdvancedRow = ({ icon, label, value, valueTooltip, progress }: AdvancedRowProps) => (
+  <Stack gap={3} style={{ minWidth: 150 }}>
+    <Group gap={6} wrap="nowrap" miw={0} justify="space-between" w="100%">
+      <Group gap={6} wrap="nowrap" miw={0}>
+        {icon}
+        <Text size="xs" c="dimmed" lineClamp={1}>
+          {label}
+        </Text>
+      </Group>
+      {valueTooltip ? (
+        <Tooltip label={valueTooltip} withArrow>
+          <Text size="xs" fw={600} lineClamp={1} style={{ maxWidth: 240 }}>
+            {value}
+          </Text>
+        </Tooltip>
+      ) : (
+        <Text size="xs" fw={600} lineClamp={1} style={{ maxWidth: 240 }}>
+          {value}
+        </Text>
+      )}
+    </Group>
+    {progress && <Progress value={progress.value} size="xs" radius="xs" color={progress.color} />}
+  </Stack>
+);
+
+const InfoRow = ({ label, value }: { label: string; value: string }) => (
+  <Group justify="space-between" wrap="nowrap" miw={0}>
+    <Text size="xs" c="dimmed" lineClamp={1}>
+      {label}
+    </Text>
+    <Text size="xs" fw={600} lineClamp={1}>
+      {value}
+    </Text>
+  </Group>
+);

packages/widgets/src/llama-cpp/index.ts

@@ -0,0 +1,24 @@
+import { IconCpu } from "@tabler/icons-react";
+
+import { getWidgetIntegrationConfig } from "@homarr/definitions";
+
+import { createWidgetDefinition } from "../definition";
+import { optionsBuilder } from "../options";
+
+export const { definition, componentLoader } = createWidgetDefinition("llamacpp", {
+  icon: IconCpu,
+  refetchInterval: 5,
+  supportsAdvancedFocus: true,
+  ...getWidgetIntegrationConfig("llamacpp"),
+  createOptions() {
+    return optionsBuilder.from((factory) => ({
+      showModelInfo: factory.switch({ defaultValue: true }),
+      showContextUsage: factory.switch({ defaultValue: true }),
+      showCacheHitRate: factory.switch({ defaultValue: true }),
+      showSpeedStats: factory.switch({ defaultValue: true }),
+      showRequests: factory.switch({ defaultValue: true }),
+      showTokenStats: factory.switch({ defaultValue: true }),
+      showSpeculative: factory.switch({ defaultValue: true }),
+    }));
+  },
+}).withDynamicImport(() => import("./component"));

packages/widgets/src/refetch-intervals.ts

@@ -14,6 +14,7 @@ export const widgetQueryRefetchIntervals = [
   },
   { queryKey: [["widget", "firewall"]], intervalSeconds: 10 },
   { queryKey: [["widget", "healthMonitoring"]], intervalSeconds: 10 },
+  { queryKey: [["widget", "llamacpp"]], intervalSeconds: 5 },
   { queryKey: [["widget", "mediaServer", "getCurrentStreams"]], intervalSeconds: 10 },
   { queryKey: [["widget", "tracearr"]], intervalSeconds: 10 },
   {

packages/widgets/src/registry.ts

@@ -59,6 +59,7 @@ import type * as video from "./video";
 import type * as vpn from "./vpn";
 import type * as weather from "./weather";
 import type * as wud from "./wud";
+import type * as llamacpp from "./llama-cpp";
 
 // Keep these imports explicit so Next.js and Turbopack can discover every widget
 // module without loading any widget definition or component eagerly.
@@ -122,6 +123,7 @@ export const widgetModuleLoaders = {
   customApi: () => import("./custom-api"),
   weather: () => import("./weather"),
   wud: () => import("./wud"),
+  llamacpp: () => import("./llama-cpp"),
 } satisfies { [TKind in WidgetKind]: () => Promise<WidgetImports[TKind]> };
 
 export type WidgetImports = {
@@ -184,6 +186,7 @@ export type WidgetImports = {
   customApi: typeof customApi;
   weather: typeof weather;
   wud: typeof wud;
+  llamacpp: typeof llamacpp;
 };
 
 type WidgetComponentLoaders = {
@@ -253,6 +256,7 @@ export const widgetComponentLoaders = {
   customApi: () => import("./custom-api/component"),
   weather: () => import("./weather/component"),
   wud: () => import("./wud/component"),
+  llamacpp: () => import("./llama-cpp/component"),
 } satisfies WidgetComponentLoaders;
 
 export type WidgetImportKey = keyof WidgetImports;