diff --git a/open-sse/providers/registry/index.js b/open-sse/providers/registry/index.js
index 4318a5a1..07b71d2c 100644
--- a/open-sse/providers/registry/index.js
+++ b/open-sse/providers/registry/index.js
@@ -125,6 +125,7 @@ import p119 from "./selfhosted-embedding.js";
 import p120 from "./fish-audio.js";
 import p121 from "./alitp-intl.js";
 import p122 from "./xquik.js";
+import p125 from "./tokenharbor.js";
 import p126 from "./dahl.js";
 import p127 from "./atria.js";
 import p129 from "./agnes.js";
@@ -254,6 +255,7 @@ export default [
   p120,
   p121,
   p122,
+  p125,
   p126,
   p127,
   p129,
diff --git a/open-sse/providers/registry/tokenharbor.js b/open-sse/providers/registry/tokenharbor.js
new file mode 100644
index 00000000..2ef04347
--- /dev/null
+++ b/open-sse/providers/registry/tokenharbor.js
@@ -0,0 +1,49 @@
+export default {
+  id: "tokenharbor",
+  priority: 120,
+  alias: "tokenharbor",
+  aliases: [
+    "th",
+    "thh",
+  ],
+  uiAlias: "tokenharbor",
+  display: {
+    name: "Token Harbor",
+    icon: "anchor",
+    color: "#0F766E",
+    textIcon: "TH",
+    website: "https://tokenharbor.ai",
+    notice: {
+      text: "OpenAI-compatible aggregator. One API key reaches every model, billed per-token from a prepaid wallet. Model ids are bare (e.g. claude-opus-5.5, gpt-6-astra, deepseek-v4.1-flash:free) and are fetched live from the provider.",
+      apiKeyUrl: "https://tokenharbor.ai/dashboard",
+    },
+  },
+  category: "apikey",
+  authType: "apikey",
+  transport: {
+    // OpenAI-compatible. `format` is left at the shared "openai" default and
+    // `thinkingFormat` is deliberately NOT declared: Token Harbor forwards
+    // requests verbatim, so each model must resolve its own thinking wire
+    // format through providers/capabilities.js. Setting a provider-wide value
+    // would force one format (e.g. claude-adaptive) onto every model.
+    baseUrl: "https://tokenharbor.ai/v1/chat/completions",
+    validateUrl: "https://tokenharbor.ai/v1/models",
+    retry: {
+      429: 2,
+    },
+  },
+  // Curated seed; the live catalogue is fetched via modelsFetcher and any other
+  // id is accepted via passthroughModels. Their catalogue rotates (the :free set
+  // in particular), so this stays deliberately small and is only the offline
+  // fallback. Ids are bare — Token Harbor does not prefix them by upstream vendor.
+  models: [
+    { id: "claude-opus-5.5", name: "Claude Opus 5.5" },
+    { id: "claude-sonnet-5", name: "Claude Sonnet 5" },
+    { id: "gpt-6-astra", name: "GPT-6 Astra" },
+    { id: "gpt-6-sol", name: "GPT-6 Sol" },
+    { id: "deepseek-v4.1-flash:free", name: "DeepSeek V4.1 Flash (Free)" },
+    { id: "grok-4.7", name: "Grok 4.7" },
+  ],
+  modelsFetcher: { url: "https://tokenharbor.ai/v1/models", type: "openai" },
+  passthroughModels: true,
+};
diff --git a/public/providers/tokenharbor.png b/public/providers/tokenharbor.png
new file mode 100644
index 00000000..7a766a64
Binary files /dev/null and b/public/providers/tokenharbor.png differ
diff --git a/src/app/api/providers/[id]/models/route.js b/src/app/api/providers/[id]/models/route.js
index 4b5f9f82..a2e09501 100644
--- a/src/app/api/providers/[id]/models/route.js
+++ b/src/app/api/providers/[id]/models/route.js
@@ -302,9 +302,8 @@ const PROVIDER_MODELS_CONFIG = {
   nvidia: createOpenAIModelsConfig("https://integrate.api.nvidia.com/v1/models"),
   assemblyai: createOpenAIModelsConfig("https://api.assemblyai.com/v1/models"),
   "vercel-ai-gateway": createOpenAIModelsConfig("https://ai-gateway.vercel.sh/v1/models"),
-  // OpenAI-compatible aggregators. All standard Bearer + OpenAI shape, so they
-  // reuse createOpenAIModelsConfig; Kira and Dahl serve their catalogue
-  // publicly, the rest need the connection API key.
+  // OpenAI-compatible aggregators.
+  tokenharbor: createOpenAIModelsConfig("https://tokenharbor.ai/v1/models"),
   dahl: createOpenAIModelsConfig("https://inference.dahl.global/v1/models"),
   atria: createOpenAIModelsConfig("https://api.atria-asi.ai/v1/models"),
   agnes: createOpenAIModelsConfig("https://apihub.agnes-ai.com/v1/models"),
diff --git a/src/app/api/providers/[id]/test/testUtils.js b/src/app/api/providers/[id]/test/testUtils.js
index 90e19d83..92873ccc 100644
--- a/src/app/api/providers/[id]/test/testUtils.js
+++ b/src/app/api/providers/[id]/test/testUtils.js
@@ -696,8 +696,7 @@ async function testApiKeyConnection(connection, effectiveProxy = null) {
         const res = await fetchWithConnectionProxy("https://api.hyperbolic.xyz/v1/models", { headers: { Authorization: `Bearer ${connection.apiKey}` } }, effectiveProxy);
         return { valid: res.ok, error: res.ok ? null : "Invalid API key" };
       }
-      // OpenAI-compatible aggregators. All standard Bearer + OpenAI shape, so the
-      // same /v1/models probe applies to each.
+      case "tokenharbor":
       case "dahl":
       case "atria":
       case "agnes":
diff --git a/tests/__baseline__/alias-baseline.json b/tests/__baseline__/alias-baseline.json
index 547f0c0e..80d3af8a 100644
--- a/tests/__baseline__/alias-baseline.json
+++ b/tests/__baseline__/alias-baseline.json
@@ -196,6 +196,7 @@
     "siliconflow": "siliconflow",
     "tencent": "hunyuan",
     "together": "together",
+    "tokenharbor": "tokenharbor",
     "tokenrouter": "tokenrouter",
     "venice": "venice",
     "vercel-ai-gateway": "vercel-ai-gateway",
@@ -299,6 +300,7 @@
     "siliconflow",
     "stability-ai",
     "together",
+    "tokenharbor",
     "tokenrouter",
     "venice",
     "vertex",
diff --git a/tests/__baseline__/providers-baseline.json b/tests/__baseline__/providers-baseline.json
index 85a3ee85..8c12ef73 100644
--- a/tests/__baseline__/providers-baseline.json
+++ b/tests/__baseline__/providers-baseline.json
@@ -1105,5 +1105,13 @@
     "baseUrl": "https://api.b.ai/v1/chat/completions",
     "validateUrl": "https://api.b.ai/v1/models",
     "format": "openai"
+  },
+  "tokenharbor": {
+    "baseUrl": "https://tokenharbor.ai/v1/chat/completions",
+    "validateUrl": "https://tokenharbor.ai/v1/models",
+    "retry": {
+      "429": 2
+    },
+    "format": "openai"
   }
 }
\ No newline at end of file
diff --git a/tests/unit/tokenharbor-provider.test.js b/tests/unit/tokenharbor-provider.test.js
new file mode 100644
index 00000000..2f4af4fa
--- /dev/null
+++ b/tests/unit/tokenharbor-provider.test.js
@@ -0,0 +1,79 @@
+import { describe, expect, it } from "vitest";
+
+import REGISTRY from "../../open-sse/providers/registry/index.js";
+import { PROVIDERS, PROVIDER_MODELS } from "../../open-sse/providers/index.js";
+import { getCapabilitiesForModel } from "../../open-sse/providers/capabilities.js";
+import { getExecutor } from "../../open-sse/executors/index.js";
+import { DefaultExecutor } from "../../open-sse/executors/default.js";
+
+describe("Token Harbor provider", () => {
+  const entry = REGISTRY.find((e) => e.id === "tokenharbor");
+
+  it("is registered as an OpenAI-compatible apikey provider", () => {
+    expect(entry).toBeDefined();
+    expect(entry.category).toBe("apikey");
+    expect(entry.authType).toBe("apikey");
+    expect(entry.alias).toBe("tokenharbor");
+    expect(entry.aliases).toContain("th");
+  });
+
+  it("points at the verified OpenAI-compatible base URL", () => {
+    expect(PROVIDERS.tokenharbor.baseUrl).toBe("https://tokenharbor.ai/v1/chat/completions");
+    expect(PROVIDERS.tokenharbor.validateUrl).toBe("https://tokenharbor.ai/v1/models");
+    // transport.format defaults to "openai" via the shared provider default
+    expect(PROVIDERS.tokenharbor.format).toBe("openai");
+  });
+
+  it("declares no provider-wide thinkingFormat so each model resolves its own", () => {
+    // Token Harbor forwards bodies verbatim. A provider-wide thinkingFormat
+    // would override capabilities.js and force one wire format (e.g.
+    // claude-adaptive) onto every model, which an OpenAI endpoint rejects.
+    expect(PROVIDERS.tokenharbor.thinkingFormat).toBeUndefined();
+  });
+
+  it("enables dynamic model discovery and passthrough", () => {
+    expect(entry.passthroughModels).toBe(true);
+    expect(entry.modelsFetcher).toMatchObject({
+      url: "https://tokenharbor.ai/v1/models",
+      type: "openai",
+    });
+  });
+
+  it("exposes a small seed of bare (unprefixed) model ids", () => {
+    const ids = (PROVIDER_MODELS.tokenharbor || []).map((m) => m.id);
+    expect(ids.length).toBeGreaterThan(0);
+    expect(ids).toContain("claude-opus-5.5");
+    // Token Harbor does not prefix ids by upstream vendor
+    expect(ids.every((id) => !id.includes("/"))).toBe(true);
+  });
+
+  it("routes through the shared DefaultExecutor (no custom adapter)", () => {
+    expect(getExecutor("tokenharbor")).toBeInstanceOf(DefaultExecutor);
+  });
+
+  it("resolves per-model capabilities from the shared tables", () => {
+    // Bare ids must still reach the canonical family patterns.
+    expect(getCapabilitiesForModel("tokenharbor", "claude-opus-5.5")).toMatchObject({
+      vision: true,
+      reasoning: true,
+      thinkingFormat: "claude-adaptive",
+    });
+    expect(getCapabilitiesForModel("tokenharbor", "gpt-6-astra")).toMatchObject({
+      reasoning: true,
+      thinkingFormat: "openai",
+    });
+  });
+
+  it("does not invent capabilities for an uncatalogued model", () => {
+    // Vision/reasoning must not be blanket-granted across the provider.
+    const caps = getCapabilitiesForModel("tokenharbor", "some-unknown-model-x");
+    expect(caps.vision).toBe(false);
+    expect(caps.reasoning).toBe(false);
+    expect(caps.thinkingFormat).toBeNull();
+  });
+
+  it("keeps every registry id unique after adding tokenharbor", () => {
+    const ids = REGISTRY.map((e) => e.id);
+    expect(new Set(ids).size).toBe(ids.length);
+  });
+});