diff --git a/docs/subsystems/providers-accounts.md b/docs/subsystems/providers-accounts.md
index 1ffe45f7..1f8cbeba 100644
--- a/docs/subsystems/providers-accounts.md
+++ b/docs/subsystems/providers-accounts.md
@@ -31,7 +31,7 @@ keys and accounts per request is in [Gateway routing and fallback](gateway-routi
 | A key's own windows | A sub2api key given a spending limit per 5 hours, day or 7 days answers its `/v1/usage` with `rate_limits` and no `remaining` (`readSub2APIKeyLimits`). With the Balance URL set to it, the key's card shows them as the windows *5 hours*, *1 day*, *7 days*, in USD, beside a balance field the user wrote; *Check balance* and `magpie provider show` give what its fullest window has left (`Balance`). `KeyAllowance` gives routing the same windows: never waited for, read in the background at most once a minute, a failed read keeping the last, and seeded from the card's last reading (`quotas.json`) after a restart.  A key bought as a plan (Kimi Code, GLM Coding Plan and Z.ai, MiniMax Coding Plan, OpenCode Go, Command Code: `planQuotaSourceOf`) gives routing its card's windows the same way (`planKeyWindows`, Zhipu's team quota behind a key with none of its own); the Usage page's read of its card is taken at once (`noteKeyAllowance`), and a key there that turns out pay-as-you-go, with no windows, is asked again only every 30 minutes (`keyNoWindowsAge`). StepFun's Step Plan is read through a sign-in, not the key, and isn't one of them. | [`sub2api_usage.go`](../../internal/provider/sub2api_usage.go), [`key_allowance.go`](../../internal/provider/key_allowance.go) |
 | Last served | `served.json` records when each account, plan or key last answered, shown as `lastServedAt` (#570). | [`served.go`](../../internal/provider/served.go) |
 | Partners | Services that pay to be listed first in the add sheet, under a heading of their own ("Partners · magpie's sponsors"). The list is `site/partners.js`, served at `usemagpie.ai/api/partners`: a change reaches users with a deploy, not a release. `Partners` never waits: it returns what was fetched last and fetches again behind the call every 6 hours (10 minutes after a failure), keeping the list in `partners.json` in the cache folder. Each entry is checked again on every read (`validPartners`): its own id, never a built-in preset's, a name, an endpoint, https only, at most 12; what only magpie's presets may set (a decision API, no key, a user endpoint) is cleared. Its logo is fetched through `FetchIcon` and kept as an uploaded icon. A partner is a preset to the rest of magpie (`Preset`, `FromPreset`, the editor), and one no longer listed still is to the providers added from it (`Seen`), without the sponsored mark. The window lists a partner in the languages it names (`langs`) only and shows its tagline in the page's language (`notes`, else `en`). `MAGPIE_PARTNERS` points it elsewhere, `off` lists none; tests list none unless they name a list. How often each listed partner was shown, opened (its row, its key page) and added is counted per UTC day (`CountPartner`, capped at 20 of each and 5 adds, kept in `partner-counts.json` in the cache folder) and sent with the next day's stats event as `magpie partner` events (`internal/stats`); nothing is counted with the stats or their usage part off. A listed partner the add sheet (or the TUI's, or `magpie providers presets`) hasn't shown yet is new (`NoticePartners` keeps those shown, as `noticed` in `partners.json`, whatever the stats); the window's add button carries a small dot for it, with its name in the title, until the sheet shows it. | [`partners.go`](../../internal/provider/partners.go), [`partner_events.go`](../../internal/provider/partner_events.go), [`site/partners.js`](../../site/partners.js) |
-| Protocol detection | Works out which of Chat, Responses and Anthropic Messages a base URL speaks when a provider is added; a Gemini URL is not detected, and the editor hides Detect while Gemini is the Base URL's API. Detection, endpoint tests and model tests share `probe`, which adds `anthropic-version` only for Anthropic Messages; OpenAI probes omit that default so a relay that refuses it can answer. User-supplied headers still apply through `Sign`. | [`detect.go`](../../internal/provider/detect.go), [`test.go`](../../internal/provider/test.go) |
+| Protocol detection | Works out which of Chat, Responses, Anthropic Messages and Gemini generateContent a base URL speaks when a provider is added. Each is asked at the URL as it takes it (`DetectBase`); Gemini's at the `…/v1beta` beside a typed `…/v1` or bare host, a Gemini model from the list picked for it, with the key in `x-goog-api-key` (#1346). Detection, endpoint tests and model tests share `probe`, which adds `anthropic-version` only for Anthropic Messages; OpenAI probes omit that default so a relay that refuses it can answer. User-supplied headers still apply through `Sign`. | [`detect.go`](../../internal/provider/detect.go), [`test.go`](../../internal/provider/test.go) |
 
 ## Runtime path
 
diff --git a/internal/gui/assets/app.js b/internal/gui/assets/app.js
index 88875420..9cee9404 100644
--- a/internal/gui/assets/app.js
+++ b/internal/gui/assets/app.js
@@ -7821,7 +7821,6 @@ function drawEditor(p, presetID) {
   // the Redaction row, shown while every address is on this computer or
   // the local network
   let showLocal = () => {};
-  let showDetect = () => {};
   if (custom) {
     name = input(draft.name, t("e.g. My Relay"));
     name.oninput = () => { draft.name = name.value; if (isNew) draft.id = slug(name.value); };
@@ -7865,7 +7864,6 @@ function drawEditor(p, presetID) {
       slide(seg, "api");
       url.placeholder = urlHint(v);
       fillEndpoints();
-      showDetect();
       showSearch();
       showLocal();
     };
@@ -7892,12 +7890,8 @@ function drawEditor(p, presetID) {
       showApi(v || draft.api);
       draft.onModelPrefs?.(); // the APIs a model can be given follow the URLs
     };
-    // Detect asks the three OpenAI and Anthropic APIs, which a Gemini
-    // API's URL has none of
-    const detect = detectAPIs(p, () => url.value, useDetected);
-    urlWrap.append(detect);
-    showDetect = () => { detect.hidden = draft.api === "gemini"; };
-    showDetect();
+    // Detect asks OpenAI's, Anthropic's and Gemini's APIs (#1346)
+    urlWrap.append(detectAPIs(p, () => url.value, useDetected));
     ed.append(...field("Base URL", urlWrap));
   }
 
@@ -9037,7 +9031,7 @@ function detectAPIs(p, base, use) {
   const box = el("div", "detect");
   const row = el("div", "detect-row");
   const go = el("button", "text action", t("Detect APIs"));
-  go.title = t("Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages) at this URL, to see which answer");
+  go.title = t("Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages, Gemini generateContent) at this URL, to see which answer");
   const each = el("button", "text action detect-each", t("Each picked model"));
   each.title = t("Ask every model picked below on each API, to see which API serves which model (up to {n}, a few at a time)", { n: DETECT_MAX });
   const model = input(draft.detectModel || "", t("model to try · empty picks one from the vendor's list"));
@@ -9067,7 +9061,7 @@ function detectAPIs(p, base, use) {
   };
   const ask = async (btn, extra) => {
     const typedBase = (base() || "").trim();
-    if (!typedBase && !(draft.chat || draft.responses || draft.anthropic || "").trim()) { status(t("Type the base URL first"), "warn"); return null; }
+    if (!typedBase && !(draft.chat || draft.responses || draft.anthropic || draft.gemini || "").trim()) { status(t("Type the base URL first"), "warn"); return null; }
     btn.classList.add("busy");
     out.hidden = false;
     out.replaceChildren(el("span", "hint", t("Asking each API…")));
@@ -9075,7 +9069,7 @@ function detectAPIs(p, base, use) {
       // the URL typed is asked as each API takes it, not as the field
       // holding it would send it; URLs given under More endpoints as they are
       const body = { ...asTyped(), id: p?.id, base: typedBase, ...extra };
-      for (const k of ["chat", "responses", "anthropic"]) if (body[k] === typedBase) body[k] = "";
+      for (const k of ["chat", "responses", "anthropic", "gemini"]) if (body[k] === typedBase) body[k] = "";
       const r = await api("provider/detect", body);
       out.replaceChildren();
       return r;
diff --git a/internal/gui/assets/i18n.js b/internal/gui/assets/i18n.js
index 416f20a1..a2fd818e 100644
--- a/internal/gui/assets/i18n.js
+++ b/internal/gui/assets/i18n.js
@@ -1712,7 +1712,7 @@ const I18N = {
     "Answered via {api} in {took}": "通过 {api} 协议在 {took}内响应",
     "{model} answered via {api} in {took}": "{model} 通过 {api} 协议在 {took}内响应",
     "{model} didn't answer via {api}: {error}": "{model} 通过 {api} 协议没有响应：{error}",
-    "Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages) at this URL, to see which answer": "向此地址的每种协议（OpenAI Chat Completions、Responses、Anthropic Messages）各发一个最小请求，检测哪些可用",
+    "Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages, Gemini generateContent) at this URL, to see which answer": "向此地址的每种协议（OpenAI Chat Completions、Responses、Anthropic Messages、Gemini generateContent）各发一个最小请求，检测哪些可用",
     "model to try · empty picks one from the vendor's list": "试用的模型 · 留空则从供应商的模型列表中挑选",
     "Type the base URL first": "请先填写 Base URL",
     "Asking each API…": "正在逐个检测协议…",
@@ -5492,7 +5492,7 @@ const I18N = {
     "Answered via {api} in {took}": "透過 {api} 協議在 {took}內回應",
     "{model} answered via {api} in {took}": "{model} 透過 {api} 協議在 {took}內回應",
     "{model} didn't answer via {api}: {error}": "{model} 透過 {api} 協議沒有回應：{error}",
-    "Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages) at this URL, to see which answer": "向此位址的每種協議（OpenAI Chat Completions、Responses、Anthropic Messages）各發一個最小請求，檢測哪些可用",
+    "Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages, Gemini generateContent) at this URL, to see which answer": "向此位址的每種協議（OpenAI Chat Completions、Responses、Anthropic Messages、Gemini generateContent）各發一個最小請求，檢測哪些可用",
     "model to try · empty picks one from the vendor's list": "試用的模型 · 留空則從供應商的模型列表中挑選",
     "Type the base URL first": "請先填寫 Base URL",
     "Asking each API…": "正在逐個檢測協議…",
@@ -9250,7 +9250,7 @@ const I18N = {
     "Answered via {api} in {took}": "{api} 経由で {took} で応答",
     "{model} answered via {api} in {took}": "{model} が {api} 経由で {took} で応答しました",
     "{model} didn't answer via {api}: {error}": "{model} は {api} 経由で応答しませんでした：{error}",
-    "Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages) at this URL, to see which answer": "この URL の各 API（OpenAI Chat Completions、Responses、Anthropic Messages）に最小のリクエストを送り、どれが応答するか調べます",
+    "Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages, Gemini generateContent) at this URL, to see which answer": "この URL の各 API（OpenAI Chat Completions、Responses、Anthropic Messages、Gemini generateContent）に最小のリクエストを送り、どれが応答するか調べます",
     "model to try · empty picks one from the vendor's list": "試すモデル · 空欄ならベンダーの一覧から選択",
     "Type the base URL first": "先にベース URL を入力してください",
     "Asking each API…": "各 API に問い合わせ中…",
@@ -13027,7 +13027,7 @@ const I18N = {
     "Answered via {api} in {took}": "Antwort über {api} in {took}",
     "{model} answered via {api} in {took}": "{model} hat über {api} in {took} geantwortet",
     "{model} didn't answer via {api}: {error}": "{model} hat über {api} nicht geantwortet: {error}",
-    "Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages) at this URL, to see which answer": "Eine minimale Anfrage an jede API (OpenAI Chat Completions, Responses, Anthropic Messages) unter dieser URL senden, um zu sehen, welche antworten",
+    "Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages, Gemini generateContent) at this URL, to see which answer": "Eine minimale Anfrage an jede API (OpenAI Chat Completions, Responses, Anthropic Messages, Gemini generateContent) unter dieser URL senden, um zu sehen, welche antworten",
     "model to try · empty picks one from the vendor's list": "zu testendes Modell · leer wählt eines aus der Modellliste des Anbieters",
     "Type the base URL first": "Geben Sie zuerst die Basis-URL ein",
     "Asking each API…": "Jede API wird angefragt…",
diff --git a/internal/gui/providers_detect_test.go b/internal/gui/providers_detect_test.go
index cb24fc45..00bca6db 100644
--- a/internal/gui/providers_detect_test.go
+++ b/internal/gui/providers_detect_test.go
@@ -28,7 +28,12 @@ func TestProviderDetect(t *testing.T) {
 	up := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		io.Copy(io.Discard, r.Body)
 		mu.Lock()
-		auths = append(auths, r.Method+" "+r.URL.Path+" "+r.Header.Get("Authorization"))
+		// Gemini's API is asked with the key in x-goog-api-key
+		auth := r.Header.Get("Authorization")
+		if g := r.Header.Get("x-goog-api-key"); g != "" {
+			auth = "goog " + g
+		}
+		auths = append(auths, r.Method+" "+r.URL.Path+" "+auth)
 		mu.Unlock()
 		switch r.URL.Path {
 		case "/v1/chat/completions":
@@ -87,7 +92,7 @@ func TestProviderDetect(t *testing.T) {
 	}
 	mu.Lock()
 	for _, a := range auths {
-		if !strings.HasSuffix(a, "Bearer sk-saved") {
+		if !strings.HasSuffix(a, "Bearer sk-saved") && !strings.HasSuffix(a, "goog sk-saved") {
 			t.Fatalf("asked %s, not with the saved key", a)
 		}
 	}
diff --git a/internal/gui/tests/provider-detect.test.cjs b/internal/gui/tests/provider-detect.test.cjs
index 3aaec7a5..993cff41 100644
--- a/internal/gui/tests/provider-detect.test.cjs
+++ b/internal/gui/tests/provider-detect.test.cjs
@@ -144,7 +144,7 @@ for (const engine of (process.env.BROWSER ? [process.env.BROWSER] : ["chromium",
       assert.deepEqual(save.body.modelPrefs, { "m-mixed": { api: "anthropic" } });
 
       const missing = await page.evaluate(() => [
-        "Detect APIs", "Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages) at this URL, to see which answer",
+        "Detect APIs", "Send the smallest request to each API (OpenAI chat completions, Responses, Anthropic messages, Gemini generateContent) at this URL, to see which answer",
         "model to try · empty picks one from the vendor's list", "Type the base URL first", "Asking each API…", "model {model}",
         "None answered: check the URL and the key, or type a model the vendor serves", "Use these", "Set the URLs of the APIs that answered; one not found there is cleared",
         "Taken · save to keep", "Ask {model} on {api} only", "Staged in Names & levels and made with the Save; Auto there gives it back", "{model} is asked on {api} once saved",
diff --git a/internal/gui/tests/provider-gemini.test.cjs b/internal/gui/tests/provider-gemini.test.cjs
index 1a1540ef..deaa74cc 100644
--- a/internal/gui/tests/provider-gemini.test.cjs
+++ b/internal/gui/tests/provider-gemini.test.cjs
@@ -1,11 +1,13 @@
 // Run with Node's test runner and Playwright on the module path; see README.md.
 // A custom provider's Base URL can be a Gemini API (#1346, NagaseMinato):
 // "Gemini compatible" beside the other APIs. A provider saved with a Gemini
-// URL opens on it, with no Detect (which asks OpenAI's and Anthropic's
-// APIs); picked for a provider that has a chat URL, the Gemini URL typed is
-// saved as its own, the chat URL kept, and the Base URL said to be Gemini's.
-// The pick fits the editor at a narrow width. In English, Chinese,
-// Japanese and German, Chromium and WebKit.
+// URL opens on it, and Detect asks a Gemini API too (NagaseMinato: 「检测」
+// 不会探测 Gemini 端点，为什么呢？): the URL typed is sent as the base, its
+// Gemini row shown with the others, and Use these keeps Gemini's URL;
+// picked for a provider that has a chat URL, the Gemini URL typed is saved
+// as its own, the chat URL kept, and the Base URL said to be Gemini's. The
+// pick and Detect's rows fit the editor at a narrow width. In English,
+// Chinese, Japanese and German, Chromium and WebKit.
 const assert = require("node:assert/strict");
 const fs = require("node:fs/promises");
 const path = require("node:path");
@@ -32,6 +34,16 @@ function serve(lang, posts) {
     if (url.pathname === "/api/state") return json(state);
     if (url.pathname === "/api/providers") return json(providers);
     if (url.pathname === "/api/groups") return json({ groups: [] });
+    if (url.pathname === "/api/provider/detect") {
+      const body = route.request().postDataJSON();
+      posts.push({ path: url.pathname, body });
+      return json({ results: [
+        { protocol: "chat", ok: false, status: 404, error: "Not Found", model: "gemini-2.5-pro", base: G },
+        { protocol: "responses", ok: false, status: 404, error: "Not Found", model: "gemini-2.5-pro", base: G },
+        { protocol: "anthropic", ok: false, status: 404, error: "Not Found", model: "gemini-2.5-pro", base: G },
+        { protocol: "gemini", ok: true, ms: 412, model: "gemini-2.5-pro", base: G },
+      ] });
+    }
     if (url.pathname.startsWith("/api/provider/")) {
       posts.push({ path: url.pathname, body: route.request().postDataJSON() });
       return json(providers);
@@ -43,10 +55,10 @@ function serve(lang, posts) {
 }
 
 const L = {
-  en: { gemini: "Gemini compatible", openai: "OpenAI compatible", save: "Save", more: "Gemini URL" },
-  zh: { gemini: "Gemini 兼容", openai: "OpenAI 兼容", save: "保存", more: "Gemini 地址" },
-  ja: { gemini: "Gemini 互換", openai: "OpenAI 互換", save: "保存", more: "Gemini URL" },
-  de: { gemini: "Gemini-kompatibel", openai: "OpenAI-kompatibel", save: "Speichern", more: "Gemini-URL" },
+  en: { gemini: "Gemini compatible", openai: "OpenAI compatible", save: "Save", more: "Gemini URL", detect: "Detect APIs", use: "Use these" },
+  zh: { gemini: "Gemini 兼容", openai: "OpenAI 兼容", save: "保存", more: "Gemini 地址", detect: "检测协议", use: "使用检测结果" },
+  ja: { gemini: "Gemini 互換", openai: "OpenAI 互換", save: "保存", more: "Gemini URL", detect: "API を検出", use: "これを使う" },
+  de: { gemini: "Gemini-kompatibel", openai: "OpenAI-kompatibel", save: "Speichern", more: "Gemini-URL", detect: "APIs erkennen", use: "Diese verwenden" },
 };
 
 for (const engine of (process.env.BROWSER ? [process.env.BROWSER] : ["chromium", "webkit"])) {
@@ -70,12 +82,33 @@ for (const engine of (process.env.BROWSER ? [process.env.BROWSER] : ["chromium",
       const on = () => ed.locator(".segs:has(.opt[data-api]) .opt.on").textContent();
       const detect = ed.locator(".detect");
 
-      // saved with a Gemini URL: opens on it, no Detect
+      // saved with a Gemini URL: opens on it, and Detect asks it
       await open("gem");
       assert.equal(await on(), S.gemini);
       assert.equal(await ed.locator(".base-url").inputValue(), G);
       assert.equal(await ed.locator(".base-url").getAttribute("placeholder"), "https://…/v1beta");
-      assert.equal(await detect.isVisible(), false, "Detect asks no Gemini API");
+      assert.equal(await detect.isVisible(), true, "Detect is there for a Gemini URL");
+      await detect.getByRole("button", { name: S.detect, exact: true }).click();
+      const gemRow = detect.locator('.detect-out .ep[data-api="gemini"]');
+      await gemRow.locator(".res.ok").waitFor();
+      const asked = posts.find((p) => p.path === "/api/provider/detect");
+      assert.equal(asked.body.base, G, "the URL typed is the base each API is asked at");
+      assert.equal(asked.body.gemini, "", "the Gemini field is the URL typed, so it is asked as Gemini takes it");
+      assert.equal(await detect.locator(".detect-out .ep").count(), 4);
+      assert.equal(await gemRow.locator(".pl").textContent(), "Gemini");
+      for (const width of [900, 440]) {
+        await page.setViewportSize({ width, height: 760 });
+        await page.waitForTimeout(100);
+        const over = await detect.evaluate((d) => {
+          const er = d.closest(".editor").getBoundingClientRect();
+          return [...d.querySelectorAll(".detect-out .ep, .detect-row > *")].map((e) => e.getBoundingClientRect().right - er.right).filter((x) => x > 1);
+        });
+        assert.deepEqual(over, [], `${width}px: Detect's rows overflow the editor`);
+      }
+      await page.setViewportSize({ width: 900, height: 760 });
+      await detect.getByRole("button", { name: S.use, exact: true }).click();
+      assert.equal(await on(), S.gemini, "Gemini answered: the Base URL stays Gemini's");
+      assert.equal(await ed.locator(".base-url").inputValue(), G);
 
       // picked for a provider with a chat URL: the Gemini URL is its own
       await open("plain");
@@ -84,7 +117,7 @@ for (const engine of (process.env.BROWSER ? [process.env.BROWSER] : ["chromium",
       await ed.locator('.segs .opt[data-api="gemini"]').click();
       assert.equal(await on(), S.gemini);
       assert.equal(await ed.locator(".base-url").inputValue(), "", "the saved chat URL stays chat's");
-      assert.equal(await detect.isVisible(), false);
+      assert.equal(await detect.isVisible(), true);
       await ed.locator("details.more > summary").click();
       const labels = await ed.locator("details.more label").allTextContents();
       assert(!labels.includes(S.more), "the Base URL's API is not under More endpoints: " + labels);
diff --git a/internal/provider/detect.go b/internal/provider/detect.go
index 84878aea..7bf98e47 100644
--- a/internal/provider/detect.go
+++ b/internal/provider/detect.go
@@ -17,32 +17,47 @@ import (
 // 请求看是否返回200): each of the three magpie speaks upstream — Chat
 // Completions, Responses and Anthropic Messages — is sent the smallest
 // request Test sends, at that URL as the API takes it, and what answered
-// is what the provider can be given. Gemini's isn't among them: magpie
-// serves it to clients but never asks a vendor on it, so a URL for it
-// would be one the provider has nowhere to keep.
+// is what the provider can be given. Since a custom provider can keep a
+// Gemini API's URL (#1346), Gemini's generateContent is asked too, at the
+// …/v1beta a relay serves it under (NagaseMinato: 「检测」不会探测 Gemini
+// 端点，为什么呢？).
+
+// detectProtocols are the APIs a detection asks, in the order it says
+// what they answered: magpie's three, then Gemini's.
+var detectProtocols = []Protocol{Chat, Responses, Anthropic, Gemini}
 
 // Detection is what one API answered at the URL it was asked at.
 type Detection struct {
 	Result
 	// Base is the URL the provider keeps for it, as the editor's fields
-	// take it: …/v1 for OpenAI's two, the root for Anthropic's.
+	// take it: …/v1 for OpenAI's two, the root for Anthropic's, …/v1beta
+	// for Gemini's.
 	Base string `json:"base"`
 }
 
 // DetectBase is base as proto takes it: an endpoint pasted whole (…/v1/chat/
-// completions) cut to its base, then the root for Anthropic, which adds
-// /v1/messages itself, and …/v1 for OpenAI's two when base is a bare host —
+// completions, …/models/gemini-2.5-pro:generateContent) cut to its base, then the root for Anthropic, which adds
+// /v1/messages itself, …/v1 for OpenAI's two when base is a bare host, and
+// for Gemini's the …/v1beta beside a …/v1 or on a bare host (GeminiBase) —
 // the editor's respellURL, in Go.
 func DetectBase(base string, proto Protocol) string {
 	u := strings.TrimRight(strings.TrimSpace(base), "/")
 	for _, end := range []string{"/chat/completions", "/responses", "/v1/messages", "/messages"} {
 		u = strings.TrimSuffix(u, end)
 	}
+	// a Gemini endpoint or a model list pasted whole
+	u = strings.TrimRight(geminiMethod.ReplaceAllString(u, ""), "/")
 	if u == "" {
 		return ""
 	}
-	if proto == Anthropic {
+	switch proto {
+	case Anthropic:
 		return strings.TrimSuffix(u, "/v1")
+	case Gemini:
+		if pu, err := url.Parse(u); err == nil && pu.Host != "" && strings.Trim(pu.Path, "/") == "v1" {
+			return u + "beta"
+		}
+		return GeminiBase(u)
 	}
 	if pu, err := url.Parse(u); err == nil && strings.Trim(pu.Path, "/") == "" {
 		return u + "/v1"
@@ -53,7 +68,7 @@ func DetectBase(base string, proto Protocol) string {
 // ErrNoURL is a detection with no URL to ask at.
 var ErrNoURL = errors.New("type the base URL first")
 
-// Detect asks each of Chat, Responses and Anthropic at the URL p has for
+// Detect asks each of Chat, Responses, Anthropic and Gemini at the URL p has for
 // it, else at base as it takes it (DetectBase), the smallest request, and
 // says what each answered, in that order. model is asked on all three;
 // with none, one from p's list is picked for each — a Claude model on
@@ -72,9 +87,9 @@ func (p Provider) Detect(ctx context.Context, base, model string) ([]Detection,
 	if model == "" {
 		ids = q.detectModels(ctx)
 	}
-	out := make([]Detection, len(Protocols))
+	out := make([]Detection, len(detectProtocols))
 	var wg sync.WaitGroup
-	for i, proto := range Protocols {
+	for i, proto := range detectProtocols {
 		m := model
 		if m == "" {
 			m = detectModel(ids, proto)
@@ -101,7 +116,7 @@ func (p Provider) detecting(base string) (Provider, error) {
 		return p, errors.New("a sign-in is asked on its agent's own API; there is nothing to detect")
 	}
 	q := p
-	for _, proto := range Protocols {
+	for _, proto := range detectProtocols {
 		u := strings.TrimSpace(p.Base(proto))
 		if u == "" {
 			u = DetectBase(base, proto)
@@ -118,6 +133,8 @@ func (p Provider) detecting(base string) (Provider, error) {
 			q.Responses = u
 		case Anthropic:
 			q.Anthropic = u
+		case Gemini:
+			q.Gemini = u
 		}
 	}
 	if len(q.Speaks()) == 0 {
@@ -197,8 +214,8 @@ const (
 )
 
 // DetectModels asks each of models (the first DetectMax, an image model
-// left out: it is asked on the images API) on each of Chat, Responses and
-// Anthropic at the URL p has for it, else at base as it takes it, and
+// left out: it is asked on the images API) on each of Chat, Responses,
+// Anthropic and Gemini at the URL p has for it, else at base as it takes it, and
 // says what each answered, model by model; and, by API, the first answer
 // from a model it served, else the first model's — what Detect says for
 // one. A probe not started before ctx ends says so, unasked.
@@ -224,8 +241,8 @@ func (p Provider) DetectModels(ctx context.Context, base string, models []string
 	sem := make(chan struct{}, DetectWide)
 	var wg sync.WaitGroup
 	for i, m := range ids {
-		out[i] = ModelDetection{Model: m, Results: make([]Detection, len(Protocols))}
-		for j, proto := range Protocols {
+		out[i] = ModelDetection{Model: m, Results: make([]Detection, len(detectProtocols))}
+		for j, proto := range detectProtocols {
 			d, ask := q.detectOne(proto, m, detectWait)
 			if ask != nil && catalog.ImagesAPI(m) {
 				d.Error, ask = "an image model: asked on the images API, not here", nil
@@ -253,8 +270,8 @@ func (p Provider) DetectModels(ctx context.Context, base string, models []string
 		}
 	}
 	wg.Wait()
-	sum := make([]Detection, len(Protocols))
-	for j := range Protocols {
+	sum := make([]Detection, len(detectProtocols))
+	for j := range detectProtocols {
 		sum[j] = out[0].Results[j]
 		for _, md := range out {
 			if md.Results[j].OK {
@@ -293,8 +310,8 @@ func (p Provider) detectModels(ctx context.Context) []string {
 }
 
 // detectModel is the model of ids to ask proto for: one made for it — a
-// Claude model on Anthropic's, an OpenAI one on Responses, any other on
-// Chat — else the first that chats.
+// Claude model on Anthropic's, an OpenAI one on Responses, a Gemini one on
+// Gemini's, any other on Chat — else the first that chats.
 func detectModel(ids []string, proto Protocol) string {
 	made := func(id string) bool {
 		switch proto {
@@ -302,6 +319,8 @@ func detectModel(ids []string, proto Protocol) string {
 			return isClaude(id)
 		case Responses:
 			return openAIModel(id)
+		case Gemini:
+			return strings.Contains(strings.ToLower(id), "gemini")
 		}
 		return !isClaude(id)
 	}
diff --git a/internal/provider/detect_test.go b/internal/provider/detect_test.go
index 89334c65..d41bbce6 100644
--- a/internal/provider/detect_test.go
+++ b/internal/provider/detect_test.go
@@ -47,7 +47,7 @@ func relay(t *testing.T) (*httptest.Server, func() []asked) {
 		json.Unmarshal(b, &body)
 		m, _ := body["model"].(string)
 		mu.Lock()
-		got = append(got, asked{r.Method, r.URL.Path, r.Header.Get("Authorization") + r.Header.Get("x-api-key"), m, body})
+		got = append(got, asked{r.Method, r.URL.Path, r.Header.Get("Authorization") + r.Header.Get("x-api-key") + r.Header.Get("x-goog-api-key"), m, body})
 		mu.Unlock()
 		switch r.Method + " " + r.URL.Path {
 		case "GET /v1/models":
@@ -81,7 +81,7 @@ func TestDetectProtocols(t *testing.T) {
 		if err != nil {
 			t.Fatal(err)
 		}
-		if len(res) != 3 {
+		if len(res) != 4 {
 			t.Fatalf("%s: %+v", typed, res)
 		}
 		want := []struct {
@@ -93,6 +93,9 @@ func TestDetectProtocols(t *testing.T) {
 			{Chat, true, srv.URL + "/v1", "gpt-5.5"},
 			{Responses, false, srv.URL + "/v1", "gpt-5.5"},
 			{Anthropic, true, srv.URL, "claude-sonnet-5"},
+			// no Gemini model listed: the first that chats, at the
+			// …/v1beta beside the …/v1, where this relay serves none
+			{Gemini, false, srv.URL + "/v1beta", "gpt-5.5"},
 		}
 		for i, w := range want {
 			r := res[i]
@@ -123,7 +126,7 @@ func TestDetectProtocols(t *testing.T) {
 			}
 		}
 		slices.Sort(posts)
-		if !slices.Equal(posts, []string{"/v1/chat/completions", "/v1/messages", "/v1/responses"}) {
+		if !slices.Equal(posts, []string{"/v1/chat/completions", "/v1/messages", "/v1/responses", "/v1beta/models/gpt-5.5:generateContent"}) {
 			t.Fatalf("%s: asked %v", typed, posts)
 		}
 	}
@@ -179,14 +182,14 @@ func TestResponsesProbesUseArrayInput(t *testing.T) {
 	}
 	t.Run("Detect", func(t *testing.T) {
 		rs, err := p.Detect(ctx, srv.URL, model)
-		if err != nil || len(rs) != len(Protocols) {
+		if err != nil || len(rs) != len(detectProtocols) {
 			t.Fatalf("Detect: %v, %+v", err, rs)
 		}
 		check(t, rs[1].Result)
 	})
 	t.Run("DetectModels", func(t *testing.T) {
 		each, sum, err := p.DetectModels(ctx, srv.URL, []string{model})
-		if err != nil || len(each) != 1 || len(each[0].Results) != len(Protocols) || len(sum) != len(Protocols) {
+		if err != nil || len(each) != 1 || len(each[0].Results) != len(detectProtocols) || len(sum) != len(detectProtocols) {
 			t.Fatalf("DetectModels: %v, %+v, %+v", err, each, sum)
 		}
 		check(t, each[0].Results[1].Result)
@@ -357,24 +360,24 @@ func TestDetectModels(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	want := map[string][3]bool{ // chat, responses, anthropic
-		"gpt-both":    {true, true, false},
-		"gpt-resp":    {false, true, false},
-		"glm-chat":    {true, false, false},
-		"claude-only": {false, false, true},
-		"gpt-image-2": {false, false, false},
+	want := map[string][4]bool{ // chat, responses, anthropic, gemini
+		"gpt-both":    {true, true, false, false},
+		"gpt-resp":    {false, true, false, false},
+		"glm-chat":    {true, false, false, false},
+		"claude-only": {false, false, true, false},
+		"gpt-image-2": {false, false, false, false},
 	}
 	if len(each) != len(want) {
 		t.Fatalf("asked %d models: %+v", len(each), each)
 	}
 	for _, md := range each {
 		w, ok := want[md.Model]
-		if !ok || len(md.Results) != 3 {
+		if !ok || len(md.Results) != 4 {
 			t.Fatalf("%+v", md)
 		}
 		for i, r := range md.Results {
-			if r.OK != w[i] || r.Model != md.Model || r.Protocol != Protocols[i] {
-				t.Fatalf("%s on %s: %+v, want ok=%v", md.Model, Protocols[i], r, w[i])
+			if r.OK != w[i] || r.Model != md.Model || r.Protocol != detectProtocols[i] {
+				t.Fatalf("%s on %s: %+v, want ok=%v", md.Model, detectProtocols[i], r, w[i])
 			}
 			if !r.OK && md.Model != "gpt-image-2" && (r.Status != 400 || !strings.Contains(r.Error, "not supported")) {
 				t.Fatalf("%s on %s said %+v", md.Model, r.Protocol, r)
@@ -384,14 +387,14 @@ func TestDetectModels(t *testing.T) {
 	if each[len(each)-1].Results[0].Error == "" {
 		t.Fatal("an image model was asked on chat")
 	}
-	if asked != 12 {
-		t.Fatalf("%d requests for 4 models on 3 APIs", asked)
+	if asked != 16 {
+		t.Fatalf("%d requests for 4 models on 4 APIs", asked)
 	}
 	for i, w := range []struct {
 		model, base string
 	}{{"gpt-both", srv.URL + "/v1"}, {"gpt-both", srv.URL + "/v1"}, {"claude-only", srv.URL}} {
 		if !sum[i].OK || sum[i].Model != w.model || sum[i].Base != w.base {
-			t.Fatalf("by API %s: %+v", Protocols[i], sum[i])
+			t.Fatalf("by API %s: %+v", detectProtocols[i], sum[i])
 		}
 	}
 
@@ -407,7 +410,7 @@ func TestDetectModels(t *testing.T) {
 	if err != nil {
 		t.Fatal(err)
 	}
-	if len(each) != DetectMax || asked != 3*DetectMax || most > DetectWide {
+	if len(each) != DetectMax || asked != 4*DetectMax || most > DetectWide {
 		t.Fatalf("%d models, %d requests, %d at once", len(each), asked, most)
 	}
 	if _, _, err := (Provider{Key: "k"}).DetectModels(context.Background(), srv.URL, []string{" "}); err == nil {
@@ -439,3 +442,82 @@ func TestDetectModelsOutOfTime(t *testing.T) {
 		t.Fatalf("%d sent past the time", n)
 	}
 }
+
+// A Gemini API is detected too (#1346, NagaseMinato: 「检测」不会探测 Gemini
+// 端点，为什么呢？): typed as the relay's …/v1, the bare host, Google's
+// …/v1beta or a generateContent URL pasted whole, it is asked at the
+// …/v1beta, a Gemini model from its own list, the key in x-goog-api-key, the
+// smallest request generateContent takes; a provider's own Gemini URL is
+// asked as it is.
+func TestDetectGemini(t *testing.T) {
+	detectHome(t)
+	var mu sync.Mutex
+	var gem []string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case r.URL.Path == "/v1beta/models" && r.Method == "GET":
+			if r.Header.Get("x-goog-api-key") != "AIza-k" {
+				http.Error(w, `{"error":{"code":401,"message":"API key not valid","status":"UNAUTHENTICATED"}}`, http.StatusUnauthorized)
+				return
+			}
+			io.WriteString(w, `{"models":[{"name":"models/text-embedding-004","supportedGenerationMethods":["embedContent"]},{"name":"models/gemini-2.5-flash","supportedGenerationMethods":["generateContent","countTokens"]}]}`)
+		case strings.HasPrefix(r.URL.Path, "/v1beta/models/") && r.Method == "POST":
+			b, _ := io.ReadAll(r.Body)
+			mu.Lock()
+			gem = append(gem, r.URL.Path+" "+r.Header.Get("x-goog-api-key")+" "+r.Header.Get("Authorization")+" "+string(b))
+			mu.Unlock()
+			if r.Header.Get("x-goog-api-key") != "AIza-k" {
+				http.Error(w, `{"error":{"code":401,"message":"API key not valid","status":"UNAUTHENTICATED"}}`, http.StatusUnauthorized)
+				return
+			}
+			io.WriteString(w, `{"candidates":[{"content":{"parts":[{"text":"Hi"}],"role":"model"},"finishReason":"MAX_TOKENS","index":0}],"usageMetadata":{"promptTokenCount":1,"candidatesTokenCount":1,"totalTokenCount":2},"modelVersion":"gemini-2.5-flash"}`)
+		default:
+			http.Error(w, `{"error":{"code":404,"message":"Not Found","status":"NOT_FOUND"}}`, http.StatusNotFound)
+		}
+	}))
+	t.Cleanup(srv.Close)
+	for _, typed := range []string{srv.URL + "/v1", srv.URL, srv.URL + "/v1beta", srv.URL + "/v1beta/models/gemini-2.5-flash:generateContent"} {
+		mu.Lock()
+		gem = nil
+		mu.Unlock()
+		res, err := Provider{Key: "AIza-k"}.Detect(context.Background(), typed, "")
+		if err != nil {
+			t.Fatal(err)
+		}
+		g := res[len(res)-1]
+		if g.Protocol != Gemini || !g.OK || g.Base != srv.URL+"/v1beta" || g.Model != "gemini-2.5-flash" {
+			t.Fatalf("%s: gemini = %+v", typed, g)
+		}
+		mu.Lock()
+		if len(gem) != 1 || gem[0] != `/v1beta/models/gemini-2.5-flash:generateContent AIza-k  {"contents":[{"role":"user","parts":[{"text":"hi"}]}],"generationConfig":{"maxOutputTokens":16}}` {
+			t.Fatalf("%s: asked %q", typed, gem)
+		}
+		mu.Unlock()
+		for _, r := range res[:len(res)-1] {
+			if r.OK {
+				t.Fatalf("%s: %s answered on a Gemini-only relay: %+v", typed, r.Protocol, r)
+			}
+		}
+	}
+
+	// the provider's own Gemini URL, and a model typed, asked as they are
+	res, err := Provider{Key: "AIza-k", Gemini: srv.URL + "/v1beta"}.Detect(context.Background(), "https://elsewhere.invalid/v1", "gemini-2.5-flash")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if g := res[3]; !g.OK || g.Base != srv.URL+"/v1beta" {
+		t.Fatalf("own URL: %+v", g)
+	}
+
+	// each model on each API: Gemini's the fourth
+	each, sum, err := Provider{Key: "AIza-k"}.DetectModels(context.Background(), srv.URL+"/v1beta", []string{"gemini-2.5-flash"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(each) != 1 || len(each[0].Results) != 4 || !each[0].Results[3].OK || !sum[3].OK || sum[3].Base != srv.URL+"/v1beta" {
+		t.Fatalf("each %+v, sum %+v", each, sum)
+	}
+	if ps := All(); len(ps) != 0 {
+		t.Fatalf("detecting saved %+v", ps)
+	}
+}
diff --git a/internal/provider/test_protocol_headers_test.go b/internal/provider/test_protocol_headers_test.go
index f5dd6a54..6093f03a 100644
--- a/internal/provider/test_protocol_headers_test.go
+++ b/internal/provider/test_protocol_headers_test.go
@@ -45,6 +45,7 @@ func TestProviderProbesUseProtocolHeaders(t *testing.T) {
 				results = p.TestModels(ctx, []string{"gpt-6-sol"})
 				want = 1
 			case "detect":
+				want = len(detectProtocols) // Gemini's too, without anthropic-version
 				detected, err := p.Detect(ctx, srv.URL, "gpt-6-sol")
 				if err != nil {
 					t.Fatal(err)
@@ -53,6 +54,7 @@ func TestProviderProbesUseProtocolHeaders(t *testing.T) {
 					results = append(results, d.Result)
 				}
 			case "detect-models":
+				want = len(detectProtocols)
 				models, _, err := p.DetectModels(ctx, srv.URL, []string{"gpt-6-sol"})
 				if err != nil {
 					t.Fatal(err)