Skip to content
Python

Azure OpenAI as its own provider card

Rohit C Prasad159a875c1e77
diff --git a/README.md b/README.md
index b5281a55..c29a8a47 100644
--- a/README.md
+++ b/README.md
@@ -87,10 +87,13 @@ Unattended runs never self-approve: their asks park in an inbox until a human an
Model access is yours: pick a provider, paste your key, switch anytime. Supported out of the box:
-**OpenAI · Anthropic · Google Gemini · BytePlus Ark · Volcengine Ark Agent Plan · Atlas Cloud · Inkling (Thinking Machines) · GLM (Z.ai) · DeepSeek · Kimi (Moonshot) · Qwen · MiniMax · Mistral · Grok (xAI) · Solar (Upstage)** - plus aggregators **OrcaRouter** and **OpenRouter**, open-weight models via **Together** and **Fireworks**, and fully local models via **Ollama**.
+**OpenAI · Azure OpenAI · Anthropic · Google Gemini · BytePlus Ark · Volcengine Ark Agent Plan · Atlas Cloud · Inkling (Thinking Machines) · GLM (Z.ai) · DeepSeek · Kimi (Moonshot) · Qwen · MiniMax · Mistral · Grok (xAI) · Solar (Upstage)** - plus aggregators **OrcaRouter** and **OpenRouter**, open-weight models via **Together** and **Fireworks**, and fully local models via **Ollama**.
A curated model list marks what we've verified for tool-calling work. Adding any model string works at your own risk.
+Azure OpenAI uses your own resource: paste its endpoint and key (or set `AZURE_OPENAI_API_KEY`), then add your deployment names as models — on Azure the model id is the deployment name.
+
+
## Privacy
OpenWorker is local-first. Everything lives on your machine: the agent loop, your conversations, connector tokens, and model keys - all in the app's local secret store. The only cloud piece is a small service that brokers OAuth handshakes for connectors. You can always use the App without signing-in - use the connectors via manually-created credentials/API-keys.
diff --git a/coworker/providers/capabilities.py b/coworker/providers/capabilities.py
index f9285fa5..c4198643 100644
--- a/coworker/providers/capabilities.py
+++ b/coworker/providers/capabilities.py
@@ -86,6 +86,15 @@ def capabilities_for(model: str) -> ModelCapabilities:
tools=True, vision=False, parallel_tool_calls=True, streaming=True
)
+ # Azure OpenAI: the id is the user's deployment name, which may say nothing about
+ # the model behind it (the GPT / o-series name checks above still apply when it
+ # does). Every current OpenAI model there calls tools in parallel; vision stays
+ # off until the name says otherwise, so images degrade to text instead of a 400.
+ if provider == "azure":
+ return ModelCapabilities(
+ tools=True, vision=False, parallel_tool_calls=True, streaming=True
+ )
+
# Conservative default for unknown models.
return ModelCapabilities(
tools=True, vision=False, parallel_tool_calls=False, streaming=True
diff --git a/coworker/providers/registry.py b/coworker/providers/registry.py
index 6dc5a6df..d66e08e6 100644
--- a/coworker/providers/registry.py
+++ b/coworker/providers/registry.py
@@ -7,8 +7,9 @@ a `ProviderClient`. The `ProviderRouter` selects a descriptor by the `provider:`
model string and builds (and caches) its client from the matching SecretStore profile.
Today: `openai` (the default — native models via the Responses API; an optional custom
-endpoint covering Azure OpenAI's `/openai/v1` and any OpenAI-compliant gateway keeps the
-Chat Completions path), `anthropic` (native Messages API via
+endpoint for vLLM and any OpenAI-compliant gateway keeps the Chat Completions path),
+`azure` (the user's own Azure OpenAI / Foundry resource through its `/openai/v1` API,
+key-authenticated; model ids are deployment names), `anthropic` (native Messages API via
`AnthropicProvider`), `gemini` (native Google GenAI API via `GeminiProvider`), `bedrock`
(models in the user's own AWS account — Claude natively, everything else via Converse),
`vertex` (the user's own GCP project — Gemini and Claude natively, open-weight via the
@@ -125,15 +126,70 @@ def _normalize_ollama_url(url: Optional[str]) -> str:
def _build_openai(profile: dict[str, Any], secrets: Any) -> ProviderClient:
# Key resolution stays in resolve_api_key (explicit → env → SecretStore), so we just
# hand over the SecretStore. Stock OpenAI (no custom endpoint) speaks the Responses
- # API — the only wire with reasoning + tools on GPT-5.6+. A custom endpoint (Azure
- # OpenAI /openai/v1, vLLM, any OpenAI-compliant gateway) keeps Chat Completions,
- # which is what compat servers implement.
+ # API — the only wire with reasoning + tools on GPT-5.6+. A custom endpoint (vLLM,
+ # any OpenAI-compliant gateway) keeps Chat Completions, which is what compat servers
+ # implement. Azure OpenAI has its own card below (`azure`); an Azure `/openai/v1`
+ # URL pasted here keeps working for users who set it up before that card existed.
base_url = ((profile or {}).get("base_url") or "").strip() or None
if base_url:
return OpenAIProvider(secrets=secrets, base_url=base_url)
return OpenAIResponsesProvider(secrets=secrets)
+AZURE_OPENAI_ENV_KEY = "AZURE_OPENAI_API_KEY"
+_AZURE_V1_SUFFIX = "/openai/v1"
+
+
+def normalize_azure_endpoint(endpoint: Optional[str]) -> str:
+ """The `/openai/v1` base URL for an Azure OpenAI or Foundry resource, from whatever the
+ user pasted: the bare resource endpoint (`https://<resource>.openai.azure.com`, also
+ `…services.ai.azure.com` and sovereign-cloud hosts), one already ending in `/openai`
+ or `/openai/v1`, or a full request URL copied from the portal (its path and
+ `?api-version=` query are dropped — the v1 API takes neither). Empty in ⇒ empty out.
+
+ Base URL per learn.microsoft.com/azure/ai-foundry/openai/api-version-lifecycle
+ (read 2026-10-26): append `/openai/v1` to the resource endpoint; no api-version.
+ """
+ raw = (endpoint or "").strip().split("?", 1)[0].split("#", 1)[0].rstrip("/")
+ if not raw:
+ return ""
+ if "://" not in raw:
+ raw = "https://" + raw
+ lowered = raw.lower()
+ if lowered.endswith(_AZURE_V1_SUFFIX):
+ return raw
+ if lowered.endswith("/openai"):
+ return raw + "/v1"
+ # A deployments URL (`…/openai/deployments/<name>/chat/completions`) or any other
+ # path under /openai: keep the origin and the /openai segment, drop the rest.
+ marker = lowered.find("/openai/")
+ if marker != -1:
+ return raw[:marker] + _AZURE_V1_SUFFIX
+ return raw + _AZURE_V1_SUFFIX
+
+
+def _build_azure(profile: dict[str, Any], secrets: Any) -> ProviderClient:
+ """Azure OpenAI through the v1 API: the stock OpenAI SDK pointed at
+ `<endpoint>/openai/v1`. The key comes from THIS provider's profile or
+ AZURE_OPENAI_API_KEY only — never the OpenAI key or SecretStore fallback, so a
+ configured OpenAI key is never sent to an Azure host (same guard as the compat
+ vendors). Chat Completions, as the OpenAI custom-endpoint path always used for Azure."""
+ p = profile or {}
+ base_url = normalize_azure_endpoint(p.get("endpoint"))
+ if not base_url:
+ raise RuntimeError(
+ "No Azure OpenAI endpoint configured — add it in Settings ▸ Models."
+ )
+ api_key = (p.get("api_key") or "").strip() or os.environ.get(
+ AZURE_OPENAI_ENV_KEY, ""
+ ).strip()
+ if not api_key:
+ raise RuntimeError(
+ "No Azure OpenAI API key configured — add it in Settings ▸ Models."
+ )
+ return OpenAIProvider(api_key=api_key, base_url=base_url)
+
+
def _build_codex(profile: dict[str, Any], secrets: Any) -> ProviderClient:
# Credentials come from the OAuth token set in the provider's own profile,
# resolved (and refreshed) at call time by the token store — never a key.
@@ -397,14 +453,46 @@ DESCRIPTORS: list[ProviderDescriptor] = [
"Custom endpoint (optional)",
secret=False,
required=False,
- placeholder="https://…/openai/v1",
- help="For Azure OpenAI, vLLM, or any OpenAI-compliant server. Leave blank for api.openai.com.",
+ placeholder="https://…/v1",
+ help="For vLLM or any OpenAI-compliant server. Leave blank for api.openai.com. "
+ "Azure OpenAI has its own card.",
),
],
build=_build_openai,
recommended_model="gpt-5.6-sol",
env_key="OPENAI_API_KEY",
),
+ # Azure OpenAI as its own card (issue #457): before it, Azure meant pasting an
+ # `/openai/v1` URL into the OpenAI card's custom endpoint, with the OpenAI key field
+ # standing in for the Azure key. Entra ID sign-in is a second auth mode to follow (#507).
+ ProviderDescriptor(
+ name="azure",
+ title="Azure OpenAI",
+ needs_key=True,
+ fields=[
+ ProviderField(
+ "endpoint",
+ "Endpoint",
+ secret=False,
+ placeholder="https://<resource>.openai.azure.com",
+ help="Your resource's endpoint, from Keys and Endpoint in the Azure portal "
+ "or the Foundry portal. /openai/v1 is added automatically; Foundry "
+ "(…services.ai.azure.com) and sovereign-cloud hosts work too.",
+ ),
+ ProviderField(
+ "api_key",
+ "Azure OpenAI API key",
+ secret=True,
+ ),
+ ],
+ build=_build_azure,
+ # No catalogue: on Azure the model id is the user's own deployment name.
+ recommended_model=None,
+ env_key=AZURE_OPENAI_ENV_KEY,
+ blurb="Your own Azure OpenAI or Foundry resource, through its v1 API — model ids "
+ "are your deployment names.",
+ ),
+
ProviderDescriptor(
name="openai-codex",
kind="subscription",
@@ -927,9 +1015,14 @@ def descriptor_configured(d: ProviderDescriptor, profile: dict[str, Any]) -> boo
return True # keyless (Ollama) — usable out of the box
profile = profile or {}
if any(f.key == "api_key" for f in d.fields):
- return bool(profile.get("api_key")) or bool(
+ has_key = bool(profile.get("api_key")) or bool(
d.env_key and os.environ.get(d.env_key)
)
+ # A key alone is not enough where the card also requires a non-secret field
+ # (Azure's endpoint): an env key with nowhere to send it is not configured.
+ return has_key and all(
+ profile.get(f.key) for f in d.fields if f.required and not f.secret
+ )
return all(profile.get(f.key) for f in d.fields if f.required)
@@ -1213,6 +1306,27 @@ def verify_provider_key(
params={"key": key},
timeout=timeout,
)
+ elif name == "azure":
+ # The Azure v1 API's own model list, at the user's resource — never
+ # api.openai.com. The key rides as a bearer token, exactly as the OpenAI SDK
+ # sends it at run time (the v1 spec takes the key in `authorization` as
+ # well as `api-key`), so a passing Test proves the path the sessions use.
+ endpoint = normalize_azure_endpoint(
+ (fields or {}).get("endpoint") or base_url
+ )
+ if not endpoint:
+ return {"ok": False, "error": "Enter the Azure OpenAI endpoint to test."}
+ resp = httpx.get(
+ endpoint + "/models",
+ headers={"Authorization": f"Bearer {key}"},
+ timeout=timeout,
+ )
+ if resp.status_code == 404:
+ return {
+ "ok": False,
+ "error": "Reached the host, but no Azure OpenAI v1 API at this "
+ "endpoint — check the resource endpoint.",
+ }
elif name == "ollama":
base = _normalize_ollama_url(base_url)
resp = httpx.get(base.rstrip("/") + "/models", timeout=timeout)
diff --git a/surfaces/gui/e2e/fixtures.ts b/surfaces/gui/e2e/fixtures.ts
index 434f1a08..8f19715c 100644
--- a/surfaces/gui/e2e/fixtures.ts
+++ b/surfaces/gui/e2e/fixtures.ts
@@ -379,7 +379,11 @@ const PROVIDERS = [
{ name: "openai", title: "OpenAI", needs_key: true, fields: [{ key: "api_key", label: "OpenAI API key", secret: true, required: true, help: "", placeholder: "sk-…" }], configured: true, values: {}, suggested_models: ["gpt-5.5"], key_set_at: "2026-06-12", last_used_at: Math.floor(Date.now() / 1000) - 7200 },
// anthropic: configured but never used ("Not used yet").
{ name: "anthropic", title: "Claude (Anthropic)", needs_key: true, fields: [{ key: "api_key", label: "API key", secret: true, required: true, help: "", placeholder: "sk-…" }], configured: true, values: {}, suggested_models: ["claude-opus-4-8"], key_set_at: null, last_used_at: null },
+ // azure: the user's own resource — a REQUIRED endpoint field (plain row, not the custom-endpoint
+ // disclosure) ahead of the key; no suggested models (ids are deployment names).
+ { name: "azure", title: "Azure OpenAI", needs_key: true, blurb: "Your own Azure OpenAI or Foundry resource, through its v1 API — model ids are your deployment names.", fields: [{ key: "endpoint", label: "Endpoint", secret: false, required: true, help: "Your resource's endpoint, from Keys and Endpoint in the Azure portal or the Foundry portal. /openai/v1 is added automatically; Foundry (…services.ai.azure.com) and sovereign-cloud hosts work too.", placeholder: "https://<resource>.openai.azure.com" }, { key: "api_key", label: "Azure OpenAI API key", secret: true, required: true, help: "", placeholder: "" }], configured: false, values: {}, suggested_models: [], key_set_at: null, last_used_at: null },
// zai: an OpenAI-compatible vendor — unconfigured, with a prefilled editable endpoint + blurb.
+
{ name: "zai", title: "Z AI (GLM)", needs_key: true, blurb: "Uses Z AI's OpenAI-compatible API — the endpoint is prefilled, just add your key.", fields: [{ key: "api_key", label: "Z AI API key", secret: true, required: true, help: "", placeholder: "" }, { key: "base_url", label: "Endpoint", secret: false, required: false, help: "Prefilled with Z AI's international endpoint.", placeholder: "https://api.z.ai/api/paas/v4", default: "https://api.z.ai/api/paas/v4" }], configured: false, values: {}, suggested_models: ["glm-5.2"], key_set_at: null, last_used_at: null },
// Ark uses two provider identities: BytePlus pay-as-you-go and Volcengine Agent Plan CN
// have independent credentials, endpoints, and strict curated model lists.
diff --git a/surfaces/gui/e2e/provider-keys.spec.ts b/surfaces/gui/e2e/provider-keys.spec.ts
index 9dc451aa..4628b555 100644
--- a/surfaces/gui/e2e/provider-keys.spec.ts
+++ b/surfaces/gui/e2e/provider-keys.spec.ts
@@ -93,3 +93,32 @@ test("the subscription provider signs in from the browser flow, no key form", as
await page.getByTestId("set-back").click();
await expect(card).toContainText("Sign in with your plan");
});
+
+test("Azure OpenAI: the endpoint is a plain required field, saved with the key", async ({
+ page,
+}) => {
+ // The Azure card (issue #457) replaces the old trick of pasting an Azure /openai/v1 URL
+ // into the OpenAI card's custom endpoint. Its endpoint is a first-class required field,
+ // rendered as an ordinary row — not behind the "Custom endpoint" disclosure.
+ await openModels(page);
+ await page.getByTestId("set-provider-azure").click();
+ await expect(page.getByTestId("set-field-endpoint")).toBeVisible();
+ await expect(page.getByTestId("set-endpoint-link")).toHaveCount(0);
+ await expect(page.getByText("model ids are your deployment names")).toBeVisible();
+ await expect(page.getByText("portal.azure.com")).toBeVisible();
+
+ await page.getByTestId("set-field-endpoint").fill("https://contoso.openai.azure.com");
+ await page.getByTestId("set-field-api_key").fill("azure-key-1234");
+ await page.getByTestId("set-test").click();
+ await expect(page.getByTestId("set-saved-pill")).toContainText("Tested & saved");
+ await expect(page.getByTestId("set-provider-azure")).toContainText("✓ Connected", {
+ timeout: 5_000,
+ });
+
+ // Reopened: the endpoint is served back from the stored values, the key never is, and
+ // the models section says what an Azure model id is.
+ await page.getByTestId("set-provider-azure").click();
+ await expect(page.getByTestId("set-field-endpoint")).toHaveValue("https://contoso.openai.azure.com");
+ await expect(page.getByTestId("set-field-api_key")).toHaveValue("");
+ await expect(page.getByTestId("set-azure-deployment-note")).toContainText("deployment name");
+});
diff --git a/surfaces/gui/e2e/settings.spec.ts b/surfaces/gui/e2e/settings.spec.ts
index 008c38c4..91ca111d 100644
--- a/surfaces/gui/e2e/settings.spec.ts
+++ b/surfaces/gui/e2e/settings.spec.ts
@@ -88,7 +88,8 @@ test("Models: provider gallery states; vendor form previews models", async ({ pa
await expect(page.getByTestId("set-provider-ollama")).toContainText("No key needed");
await expect(page.getByTestId("provider-group-local")).toContainText("Ollama");
await expect(page.getByTestId("set-provider-zai")).not.toHaveClass(/bg-panel/);
- // Five API-key providers fit under the fold; the "Show N more" line only appears past six.
+ // Six API-key providers fit under the fold; the "Show N more" line only appears past six.
+
await expect(page.getByTestId("provider-more-api_key")).toHaveCount(0);
await expect(page.getByTestId("pick-model-btn")).toBeVisible();
await expect(page.getByTestId("provider-search")).toBeVisible();
diff --git a/surfaces/gui/src/components/ManageTabs.tsx b/surfaces/gui/src/components/ManageTabs.tsx
index 920b7f68..a666875b 100644
--- a/surfaces/gui/src/components/ManageTabs.tsx
+++ b/surfaces/gui/src/components/ManageTabs.tsx
@@ -197,7 +197,14 @@ export function ModelsTab() {
<p className="text-meta text-muted mb-2.5 leading-relaxed">
{t("manage.models_help")}
</p>
+ {ps.sel === "azure" && (
+ // No catalogue on Azure: the id to add is the deployment's name.
+ <p className="text-meta text-muted mb-2.5 leading-relaxed" data-testid="set-azure-deployment-note">
+ {t("manage.azure_deployment_help")}
+ </p>
+ )}
<ModelChecklist
+
provider={ps.sel}
knownProviders={knownNames}
suggested={info?.suggested_models || []}
diff --git a/surfaces/gui/src/locales/en.json b/surfaces/gui/src/locales/en.json
index c1fecfe0..c023fb94 100644
--- a/surfaces/gui/src/locales/en.json
+++ b/surfaces/gui/src/locales/en.json
@@ -1326,6 +1326,8 @@
"mcp_waiting_browser": "waiting for browser…",
"models": "Models",
"models_help": "Ticked models show in the composer's picker; the black badge marks the default for new sessions.",
+ "azure_deployment_help": "On Azure the model id is your deployment name — add each deployment you want to use, as named in the Azure or Foundry portal.",
+
"one_click_coming": "One-click sign-in is coming soon — connect manually below for now:",
"optional": "(optional)",
"or_connect_manually": "or connect manually:",
diff --git a/surfaces/gui/src/locales/zh.json b/surfaces/gui/src/locales/zh.json
index fa111121..26cbd7f0 100644
--- a/surfaces/gui/src/locales/zh.json
+++ b/surfaces/gui/src/locales/zh.json
@@ -1309,6 +1309,8 @@
"mcp_waiting_browser": "正在等待浏览器…",
"models": "模型",
"models_help": "已勾选的模型会显示在输入框的选择器中;黑色徽标表示新会话的默认模型。",
+ "azure_deployment_help": "在 Azure 上,模型 ID 就是你的部署名称——按 Azure 或 Foundry 门户中的名称,逐个添加你要使用的部署。",
+
"one_click_coming": "一键登录即将推出 —— 目前请在下方手动连接:",
"optional": "(可选)",
"or_connect_manually": "或手动连接:",
diff --git a/surfaces/gui/src/providers/ProviderSetup.test.tsx b/surfaces/gui/src/providers/ProviderSetup.test.tsx
index 3ac4ed04..34538812 100644
--- a/surfaces/gui/src/providers/ProviderSetup.test.tsx
+++ b/surfaces/gui/src/providers/ProviderSetup.test.tsx
@@ -39,13 +39,35 @@ const BEDROCK: ProviderInfo = {
],
};
-function makePs(fields: Record<string, string>, setFieldValue = vi.fn()): ProviderSetupState {
+// Azure OpenAI: a REQUIRED non-secret endpoint ahead of the key — a plain field row, since it
+// is keyed `endpoint`, not `base_url` (which the form folds behind the custom-endpoint link).
+const AZURE: ProviderInfo = {
+ name: "azure",
+ title: "Azure OpenAI",
+ needs_key: true,
+ configured: false,
+ values: {},
+ suggested_models: [],
+ recommended_model: null,
+ blurb: "Your own Azure OpenAI or Foundry resource, through its v1 API — model ids are your deployment names.",
+ fields: [
+ { key: "endpoint", label: "Endpoint", secret: false, required: true, help: "/openai/v1 is added automatically.", placeholder: "https://<resource>.openai.azure.com" },
+ { key: "api_key", label: "Azure OpenAI API key", secret: true, required: true, help: "", placeholder: "" },
+ ],
+};
+
+function makePs(
+ fields: Record<string, string>,
+ setFieldValue = vi.fn(),
+ info: ProviderInfo = BEDROCK,
+): ProviderSetupState {
return {
- providers: [BEDROCK],
- ordered: [BEDROCK],
+ providers: [info],
+ ordered: [info],
refreshProviders: async () => {},
- sel: "bedrock",
- info: BEDROCK,
+ sel: info.name,
+ info,
+
fields,
setFieldValue,
dirty: false,
@@ -136,3 +158,24 @@ describe("Atlas Cloud provider presentation", () => {
expect(container.querySelector("img")).not.toBeNull();
});
});
+
+describe("Azure OpenAI provider presentation", () => {
+ it("renders the endpoint as a plain required row, not the custom-endpoint disclosure", () => {
+ render(<ProviderForm ps={makePs({}, vi.fn(), AZURE)} tp="t" />);
+ const endpoint = screen.getByTestId("t-field-endpoint") as HTMLInputElement;
+ expect(endpoint.type).toBe("text");
+ expect(endpoint.placeholder).toBe("https://<resource>.openai.azure.com");
+ expect(screen.queryByTestId("t-endpoint-link")).toBeNull();
+ // Test sits beside the key (the required secret), as on every keyed card.
+ expect(screen.getByTestId("t-field-api_key")).toBeTruthy();
+ expect(screen.getByTestId("t-test")).toBeTruthy();
+ expect(screen.getByText(/deployment names/)).toBeTruthy();
+ });
+
+ it("links to the Azure portal for a key and wears the bundled mark", () => {
+ expect(KEY_HELP.azure.label).toBe("portal.azure.com");
+ expect(KEY_HELP.azure.url).toMatch(/^https:\/\/portal\.azure\.com\//);
+ const { container } = render(<ProviderMark name="azure" title="Azure OpenAI" />);
+ expect(container.querySelector("img")).not.toBeNull();
+ });
+});
diff --git a/surfaces/gui/src/providers/ProviderSetup.tsx b/surfaces/gui/src/providers/ProviderSetup.tsx
index 1227b51f..8469cb34 100644
--- a/surfaces/gui/src/providers/ProviderSetup.tsx
+++ b/surfaces/gui/src/providers/ProviderSetup.tsx
@@ -34,6 +34,9 @@ const LOCAL_HELP: Record<string, { desc: string; link: string; url: string }> =
export const KEY_HELP: Record<string, { url: string; label: string; labelKey?: string }> = {
anthropic: { url: "https://console.anthropic.com/settings/keys", label: "console.anthropic.com" },
openai: { url: "https://platform.openai.com/api-keys", label: "platform.openai.com" },
+ // The Azure portal's Azure OpenAI hub: pick the resource, then Keys and Endpoint.
+ azure: { url: "https://portal.azure.com/#view/Microsoft_Azure_ProjectOxford/CognitiveServicesHub/~/OpenAI", label: "portal.azure.com" },
+
gemini: { url: "https://aistudio.google.com/apikey", label: "aistudio.google.com" },
ark: { url: "https://console.byteplus.com/ark/region:ark+ap-southeast-1/apiKey", label: "console.byteplus.com" },
"ark-agent-plan-cn": { url: "https://console.volcengine.com/ark/region:cn-beijing/openManagement?LLM=%7B%7D&advancedActiveKey=agentPlan", label: "console.volcengine.com" },
diff --git a/surfaces/gui/src/providers/logos.ts b/surfaces/gui/src/providers/logos.ts
index 7e2995e9..ddb0cd95 100644
--- a/surfaces/gui/src/providers/logos.ts
+++ b/surfaces/gui/src/providers/logos.ts
@@ -7,6 +7,9 @@
import anthropic from "./logos/anthropic.svg";
import openai from "./logos/openai.svg";
+// Azure: a simplified "A" mark drawn in-house in Azure blues (no vendored asset).
+import azure from "./logos/azure.svg";
+
import gemini from "./logos/gemini.svg";
import byteplus from "./logos/byteplus.svg";
import volcengine from "./logos/volcengine.svg";
@@ -44,7 +47,9 @@ export const PROVIDER_LOGOS: Record<string, string> = {
// The subscription provider wears the same vendor mark — it's the same models,
// different billing (owner call 2026-08-21: no bare-letter monogram).
"openai-codex": openai,
+ azure,
gemini,
+
ark: byteplus,
"ark-agent-plan-cn": volcengine,
meta,
@@ -79,8 +84,10 @@ export const PROVIDER_LOGOS: Record<string, string> = {
export const PROVIDER_ORDER = [
"anthropic",
"openai",
+ "azure",
"gemini",
"ark",
+
"ark-agent-plan-cn",
"meta",
"ollama",
diff --git a/surfaces/gui/src/providers/logos/azure.svg b/surfaces/gui/src/providers/logos/azure.svg
new file mode 100644
index 00000000..f8228d1b
--- /dev/null
+++ b/surfaces/gui/src/providers/logos/azure.svg
@@ -0,0 +1 @@
+<svg height="1em" style="flex:none;line-height:1" viewBox="0 0 24 24" width="1em" xmlns="http://www.w3.org/2000/svg"><title>Azure OpenAI</title><path d="M8.3 2.5h5.5L8 20.2a1.1 1.1 0 01-1.05.76H2.35a1.1 1.1 0 01-1.04-1.46L7.25 3.26A1.1 1.1 0 018.3 2.5z" fill="#1E8BE0"/><path d="M15.2 2.5h.95c.47 0 .89.3 1.04.75l5.46 16.25a1.1 1.1 0 01-1.04 1.46H11.3l7.34-2.6-4.9-5.86L15.2 2.5z" fill="#1E8BE0" fill-opacity=".55"/><path d="M14 9.24l4.64 5.12-7.34 2.6-4.5 3.26 9.13-.06c.56 0 .9-.6.63-1.08L14 9.24z" fill="#0F62B8"/></svg>
diff --git a/tests/test_provider_router.py b/tests/test_provider_router.py
index ccf9bdd8..d399566d 100644
--- a/tests/test_provider_router.py
+++ b/tests/test_provider_router.py
@@ -653,3 +653,49 @@ def test_manager_key_hygiene_stamps(tmp_path, monkeypatch):
mgr2 = SessionManager(data_dir=tmp_path)
provs2 = {p["name"]: p for p in mgr2.get_providers()}
assert provs2["deepseek"]["last_used_at"] == first
+
+
+def test_manager_azure_card_saves_endpoint_and_key_and_verifies_at_the_resource(
+ tmp_path, monkeypatch
+):
+ """Settings ▸ Models for Azure OpenAI: both fields are required to save; the stored
+ endpoint comes back for prefilling, the key never does; Test probes the resource with
+ the merged (stored + typed) fields; nothing is auto-added to the picker (no catalogue)."""
+ monkeypatch.setenv("COWORKER_STATE_DIR", str(tmp_path / "state"))
+ monkeypatch.delenv("AZURE_OPENAI_API_KEY", raising=False)
+ from coworker.server.manager import SessionManager
+
+ mgr = SessionManager(data_dir=tmp_path)
+ res = mgr.set_provider("azure", {"api_key": "az-key"})
+ assert res["ok"] is False and "Endpoint" in res["error"]
+ res = mgr.set_provider(
+ "azure", {"endpoint": "https://contoso.openai.azure.com", "api_key": "az-key"}
+ )
+ assert res["ok"] and res["recommended_model"] is None
+
+ row = next(p for p in mgr.get_providers() if p["name"] == "azure")
+ assert row["configured"] is True
+ assert row["values"] == {"endpoint": "https://contoso.openai.azure.com"}
+ assert row["suggested_models"] == []
+ assert "api_key" not in row["values"]
+ assert not any(m.startswith("azure:") for m in mgr.get_settings()["models"])
+
+ # Test with the stored credentials: the probe goes to the resource's v1 model list.
+ seen: dict = {}
+
+ def fake_get(url, **kwargs):
+ seen["url"] = url
+ seen.update(kwargs)
+ from types import SimpleNamespace
+
+ return SimpleNamespace(status_code=200)
+
+ monkeypatch.setattr("httpx.get", fake_get)
+ assert mgr.verify_provider("azure", {}) == {"ok": True}
+ assert seen["url"] == "https://contoso.openai.azure.com/openai/v1/models"
+ assert seen["headers"]["Authorization"] == "Bearer az-key"
+
+ # A deployment the user adds routes to the azure provider.
+ added = mgr.add_model("azure:gpt-5-prod")
+ assert added["ok"] and "azure:gpt-5-prod" in added["models"]
+ assert mgr._model_provider("azure:gpt-5-prod") == "azure"
diff --git a/tests/test_provider_verify.py b/tests/test_provider_verify.py
index c16fd857..2f990fa4 100644
--- a/tests/test_provider_verify.py
+++ b/tests/test_provider_verify.py
@@ -314,3 +314,46 @@ def test_verify_silent_exception_falls_back_to_class_name_only(monkeypatch):
_patch_get(monkeypatch, raise_exc=Weird())
assert verify_provider_key("openai", api_key="sk-x") == {"ok": False, "error": "Couldn't reach OpenAI (Weird)."}
+
+
+# -- verify_provider_key: Azure OpenAI probes the user's resource, never api.openai.com ----
+def test_verify_azure_lists_models_at_the_resource_v1_api(monkeypatch):
+ cap: dict = {}
+ _patch_get(monkeypatch, status=200, capture=cap)
+ assert verify_provider_key(
+ "azure", api_key="az-key", fields={"endpoint": "https://contoso.openai.azure.com"}
+ ) == {"ok": True}
+ assert cap["url"] == "https://contoso.openai.azure.com/openai/v1/models"
+ assert cap["headers"]["Authorization"] == "Bearer az-key"
+ # One probe: /openai/v1 is always normalised in, so no completions-route check follows.
+ assert [c["url"] for c in cap["calls"]] == ["https://contoso.openai.azure.com/openai/v1/models"]
+ assert not any("api.openai.com" in c["url"] for c in cap["calls"])
+
+
+def test_verify_azure_accepts_an_endpoint_already_on_v1(monkeypatch):
+ cap: dict = {}
+ _patch_get(monkeypatch, status=200, capture=cap)
+ verify_provider_key(
+ "azure", api_key="az-key", fields={"endpoint": "https://contoso.openai.azure.com/openai/v1/"}
+ )
+ assert cap["url"] == "https://contoso.openai.azure.com/openai/v1/models"
+
+
+def test_verify_azure_without_an_endpoint_asks_for_one_without_a_request(monkeypatch):
+ cap: dict = {}
+ _patch_get(monkeypatch, status=200, capture=cap)
+ res = verify_provider_key("azure", api_key="az-key", fields={})
+ assert res["ok"] is False and "endpoint" in res["error"].lower()
+ assert "calls" not in cap
+
+
+def test_verify_azure_bad_key_and_wrong_host_read_differently(monkeypatch):
+ _patch_get(monkeypatch, status=401)
+ assert verify_provider_key(
+ "azure", api_key="az-bad", fields={"endpoint": "https://contoso.openai.azure.com"}
+ ) == {"ok": False, "error": "Invalid API key."}
+ _patch_get(monkeypatch, status=404)
+ res = verify_provider_key(
+ "azure", api_key="az-key", fields={"endpoint": "https://example.com"}
+ )
+ assert res["ok"] is False and "no Azure OpenAI v1 API" in res["error"]
diff --git a/tests/test_providers.py b/tests/test_providers.py
index 6785471b..d7e87b4a 100644
--- a/tests/test_providers.py
+++ b/tests/test_providers.py
@@ -5,6 +5,9 @@ from __future__ import annotations
import json
from types import SimpleNamespace
+import pytest
+
+
from coworker.providers import (
AssistantTurn,
ModelCapabilities,
@@ -680,3 +683,132 @@ def test_over_limit_max_tokens_is_dropped_and_retried():
calls = client.chat.completions.calls
assert turn.text == "ok" and len(calls) == 2
assert "max_tokens" not in calls[1]
+
+
+# -- Azure OpenAI (its own card, issue #457) ---------------------------------------------
+# Verified against learn.microsoft.com/azure/ai-foundry/openai/api-version-lifecycle
+# (read 2026-10-26): base URL = resource endpoint + `/openai/v1`, no api-version, the stock
+# OpenAI client with an API key; the model id is the deployment name.
+
+
+def test_azure_descriptor_shape():
+ from coworker.providers.registry import get_descriptor
+
+ d = get_descriptor("azure")
+ assert d is not None and d.needs_key and d.kind == "api_key"
+ assert d.title == "Azure OpenAI"
+ assert [f.key for f in d.fields] == ["endpoint", "api_key"]
+ endpoint, key = d.fields
+ # The endpoint is a REQUIRED plain field — keyed `endpoint`, not `base_url`, so the GUI
+ # renders it as an ordinary row instead of folding it behind the custom-endpoint link.
+ assert endpoint.required and not endpoint.secret
+ assert key.required and key.secret
+ assert d.env_key == "AZURE_OPENAI_API_KEY"
+ assert d.recommended_model is None # no catalogue: ids are deployment names
+ assert "deployment" in d.blurb
+
+
+ "pasted,expected",
+ [
+ ("https://contoso.openai.azure.com", "https://contoso.openai.azure.com/openai/v1"),
+ ("https://contoso.openai.azure.com/", "https://contoso.openai.azure.com/openai/v1"),
+ (" https://contoso.openai.azure.com/openai/v1/ ", "https://contoso.openai.azure.com/openai/v1"),
+ ("https://contoso.openai.azure.com/openai", "https://contoso.openai.azure.com/openai/v1"),
+ # Foundry and sovereign-cloud hosts are not second-guessed — any https host goes.
+ ("https://contoso.services.ai.azure.com", "https://contoso.services.ai.azure.com/openai/v1"),
+ ("https://contoso.openai.azure.us/", "https://contoso.openai.azure.us/openai/v1"),
+ ("contoso.openai.azure.com", "https://contoso.openai.azure.com/openai/v1"),
+ # A request URL copied from the portal: path and api-version query dropped.
+ (
+ "https://contoso.openai.azure.com/openai/deployments/gpt-5/chat/completions?api-version=2024-10-21",
+ "https://contoso.openai.azure.com/openai/v1",
+ ),
+ ("", ""),
+ (None, ""),
+ ],
+)
+def test_azure_endpoint_normalisation(pasted, expected):
+ from coworker.providers.registry import normalize_azure_endpoint
+
+ assert normalize_azure_endpoint(pasted) == expected
+
+
+def test_azure_builder_targets_the_v1_api_with_the_profile_key():
+ from coworker.providers.openai_provider import OpenAIProvider
+ from coworker.providers.registry import build_provider_client
+
+ p = build_provider_client(
+ "azure", {"endpoint": "https://contoso.openai.azure.com", "api_key": "az-key"}, None
+ )
+ assert isinstance(p, OpenAIProvider) # Chat Completions, as the custom-endpoint path did
+ assert p._base_url == "https://contoso.openai.azure.com/openai/v1"
+ assert p._api_key == "az-key"
+
+
+def test_azure_builder_env_key_fallback(monkeypatch):
+ from coworker.providers.registry import build_provider_client
+
+ monkeypatch.setenv("AZURE_OPENAI_API_KEY", "az-env")
+ p = build_provider_client("azure", {"endpoint": "https://contoso.openai.azure.com/openai/v1"}, None)
+ assert p._api_key == "az-env"
+
+
+def test_azure_builder_never_leaks_the_openai_key(monkeypatch):
+ """Same guard as the compat vendors: an OpenAI key (env or SecretStore) is never sent
+ to an Azure host — a missing Azure key fails fast, naming Azure."""
+ from coworker.providers.registry import build_provider_client
+
+ class Store:
+ def get(self, name):
+ return {"api_key": "sk-openai-stored"} if name == "provider:openai" else None
+
+ monkeypatch.setenv("OPENAI_API_KEY", "sk-openai-real")
+ monkeypatch.delenv("AZURE_OPENAI_API_KEY", raising=False)
+ with pytest.raises(RuntimeError, match="Azure OpenAI API key"):
+ build_provider_client("azure", {"endpoint": "https://contoso.openai.azure.com"}, Store())
+
+
+def test_azure_builder_requires_an_endpoint(monkeypatch):
+ from coworker.providers.registry import build_provider_client
+
+ with pytest.raises(RuntimeError, match="endpoint"):
+ build_provider_client("azure", {"api_key": "az-key"}, None)
+
+
+def test_azure_configured_needs_endpoint_and_key(monkeypatch):
+ """An env key alone is not a working setup — there is nowhere to send it."""
+ from coworker.providers.registry import descriptor_configured, get_descriptor
+
+ d = get_descriptor("azure")
+ monkeypatch.delenv("AZURE_OPENAI_API_KEY", raising=False)
+ assert not descriptor_configured(d, {})
+ assert not descriptor_configured(d, {"api_key": "k"})
+ assert not descriptor_configured(d, {"endpoint": "https://c.openai.azure.com"})
+ assert descriptor_configured(d, {"endpoint": "https://c.openai.azure.com", "api_key": "k"})
+ monkeypatch.setenv("AZURE_OPENAI_API_KEY", "env")
+ assert not descriptor_configured(d, {})
+ assert descriptor_configured(d, {"endpoint": "https://c.openai.azure.com"})
+
+
+def test_azure_models_route_by_prefix_and_default_to_agentic_capabilities():
+ from coworker.providers.capabilities import capabilities_for
+ from coworker.providers.router import ProviderRouter
+
+ router = ProviderRouter(secrets=None)
+ assert router._provider_name("azure:my-gpt-deployment") == "azure"
+ assert ProviderRouter._bare("azure:my-gpt-deployment") == "my-gpt-deployment"
+ # An arbitrary deployment name: tools + parallel calls, no vision claimed.
+ caps = capabilities_for("azure:my-gpt-deployment")
+ assert caps.tools and caps.parallel_tool_calls and caps.streaming and not caps.vision
+ # A deployment named after the model keeps the model's own profile.
+ assert capabilities_for("azure:gpt-5.5").vision
+ assert not capabilities_for("azure:o4-mini").parallel_tool_calls
+
+
+def test_openai_custom_endpoint_help_points_azure_users_at_the_card():
+ from coworker.providers.registry import get_descriptor
+
+ base = next(f for f in get_descriptor("openai").fields if f.key == "base_url")
+ assert "Azure OpenAI has its own card" in base.help
+ assert not base.required # the old Azure-URL-in-OpenAI setup keeps working