feat: add vertex-anthropic provider

apps/gateway/src/chat/chat.ts

@@ -572,7 +572,12 @@ function usesGoogleQueryToken(provider: string): boolean {
 }
 
 function isGoogleCompatibleProvider(provider: string): boolean {
-	return usesGoogleQueryToken(provider);
+	return (
+		provider === "google-ai-studio" ||
+		provider === "glacier" ||
+		provider === "google-vertex" ||
+		provider === "quartz"
+	);
 }
 
 // Pre-compiled regex pattern to avoid recompilation per request
@@ -3916,7 +3921,7 @@ chat.openapi(completions, async (c) => {
 	}
 
 	// Anthropic does not allow temperature and top_p to be set simultaneously
-	if (usedProvider === "anthropic") {
+	if (usedProvider === "anthropic" || usedProvider === "vertex-anthropic") {
 		if (temperature !== undefined && top_p !== undefined) {
 			top_p = undefined;
 		}
@@ -5707,7 +5712,8 @@ chat.openapi(completions, async (c) => {
 				const shouldBufferForHealing =
 					streamingIsJsonResponseFormat &&
 					(streamingResponseHealingEnabled === true ||
-						(usedProvider === "anthropic" &&
+						((usedProvider === "anthropic" ||
+							usedProvider === "vertex-anthropic") &&
 							response_format?.type === "json_object") ||
 						(usedProvider === "aws-bedrock" &&
 							response_format?.type === "json_object") ||
@@ -6489,7 +6495,11 @@ chat.openapi(completions, async (c) => {
 								}
 
 								// For Anthropic, if we have partial usage data, complete it
-								if (usedProvider === "anthropic" && transformedData.usage) {
+								if (
+									(usedProvider === "anthropic" ||
+										usedProvider === "vertex-anthropic") &&
+									transformedData.usage
+								) {
 									const usage = transformedData.usage;
 									if (
 										usage.output_tokens !== undefined &&
@@ -6559,7 +6569,10 @@ chat.openapi(completions, async (c) => {
 
 								// For Anthropic streaming tool calls, enrich delta chunks with id/type/name
 								// from the initial content_block_start event. This ensures OpenAI SDK compatibility.
-								if (usedProvider === "anthropic") {
+								if (
+									usedProvider === "anthropic" ||
+									usedProvider === "vertex-anthropic"
+								) {
 									const toolCalls =
 										transformedData.choices?.[0]?.delta?.tool_calls;
 									if (toolCalls && toolCalls.length > 0) {
@@ -6686,7 +6699,8 @@ chat.openapi(completions, async (c) => {
 								// use raw data. For others (like aws-bedrock), use transformed OpenAI format.
 								const contentChunk = extractContent(
 									isGoogleCompatibleProvider(usedProvider) ||
-										usedProvider === "anthropic"
+										usedProvider === "anthropic" ||
+										usedProvider === "vertex-anthropic"
 										? data
 										: transformedData,
 									usedProvider,
@@ -6717,7 +6731,10 @@ chat.openapi(completions, async (c) => {
 
 								// Track web search calls for cost calculation
 								// Check for web search results based on provider-specific data
-								if (usedProvider === "anthropic") {
+								if (
+									usedProvider === "anthropic" ||
+									usedProvider === "vertex-anthropic"
+								) {
 									// For Anthropic, count web_search_tool_result blocks
 									if (
 										data.type === "content_block_start" &&
@@ -6758,7 +6775,8 @@ chat.openapi(completions, async (c) => {
 								// use raw data. For others, use transformed OpenAI format.
 								const reasoningContentChunk = extractReasoning(
 									isGoogleCompatibleProvider(usedProvider) ||
-										usedProvider === "anthropic"
+										usedProvider === "anthropic" ||
+										usedProvider === "vertex-anthropic"
 										? data
 										: transformedData,
 									usedProvider,
@@ -6786,7 +6804,8 @@ chat.openapi(completions, async (c) => {
 
 										// For Anthropic content_block_delta events, match by content block index
 										if (
-											usedProvider === "anthropic" &&
+											(usedProvider === "anthropic" ||
+												usedProvider === "vertex-anthropic") &&
 											newCall._contentBlockIndex !== undefined
 										) {
 											existingCall =
@@ -6831,6 +6850,7 @@ chat.openapi(completions, async (c) => {
 										}
 										break;
 									case "anthropic":
+									case "vertex-anthropic":
 										if (
 											data.type === "message_delta" &&
 											data.delta?.stop_reason
@@ -9226,7 +9246,7 @@ chat.openapi(completions, async (c) => {
 	const shouldHealNonStreaming =
 		isJsonResponseFormat &&
 		(responseHealingEnabled === true ||
-			(usedProvider === "anthropic" &&
+			((usedProvider === "anthropic" || usedProvider === "vertex-anthropic") &&
 				response_format?.type === "json_object") ||
 			(usedProvider === "aws-bedrock" &&
 				response_format?.type === "json_object") ||

apps/gateway/src/chat/tools/extract-content.ts

@@ -14,6 +14,7 @@ export function extractContent(data: any, provider: Provider): string {
 			return contentParts.map((part: any) => part.text).join("") ?? "";
 		}
 		case "anthropic":
+		case "vertex-anthropic":
 			if (data.type === "content_block_delta" && data.delta?.text) {
 				return data.delta.text;
 			} else if (data.delta?.text) {

apps/gateway/src/chat/tools/extract-reasoning.ts

@@ -2,39 +2,68 @@ import { extractReasoningDetailsText } from "./reasoning-details.js";
 
 import type { Provider } from "@llmgateway/models";
 
-/**
- * Extracts reasoning content from streaming data based on provider format
- */
-export function extractReasoning(data: any, provider: Provider): string {
+interface AnthropicStreamChunk {
+	type?: string;
+	delta?: { type?: string; thinking?: string };
+}
+
+interface GoogleStreamChunk {
+	candidates?: Array<{
+		content?: { parts?: Array<{ thought?: boolean; text?: string }> };
+	}>;
+}
+
+interface OpenAIStreamChunk {
+	choices?: Array<{
+		delta?: {
+			reasoning?: string;
+			reasoning_content?: string;
+			reasoning_details?: unknown;
+		};
+	}>;
+}
+
+type StreamingChunk =
+	| AnthropicStreamChunk
+	| GoogleStreamChunk
+	| OpenAIStreamChunk;
+
+export function extractReasoning(
+	data: StreamingChunk,
+	provider: Provider,
+): string {
 	switch (provider) {
-		case "anthropic": {
-			// Handle Anthropic thinking content blocks in streaming format
+		case "anthropic":
+		case "vertex-anthropic": {
+			const chunk = data as AnthropicStreamChunk;
 			if (
-				data.type === "content_block_delta" &&
-				data.delta?.type === "thinking_delta" &&
-				data.delta?.thinking
+				chunk.type === "content_block_delta" &&
+				chunk.delta?.type === "thinking_delta" &&
+				chunk.delta?.thinking
 			) {
-				// This is a thinking delta - return the thinking content
-				return data.delta.thinking;
+				return chunk.delta.thinking;
 			}
 			return "";
 		}
 		case "google-ai-studio":
 		case "glacier":
 		case "google-vertex":
 		case "quartz": {
-			const parts = data.candidates?.[0]?.content?.parts ?? [];
-			const reasoningParts = parts.filter((part: any) => part.thought);
-			return reasoningParts.map((part: any) => part.text).join("") ?? "";
+			const chunk = data as GoogleStreamChunk;
+			const parts = chunk.candidates?.[0]?.content?.parts ?? [];
+			const reasoningParts = parts.filter((part) => part.thought);
+			return reasoningParts.map((part) => part.text).join("") ?? "";
 		}
-		default: // OpenAI format (includes GLM/ZAI which use reasoning_content)
+		default: {
+			const chunk = data as OpenAIStreamChunk;
 			return (
-				data.choices?.[0]?.delta?.reasoning ??
-				data.choices?.[0]?.delta?.reasoning_content ??
+				chunk.choices?.[0]?.delta?.reasoning ??
+				chunk.choices?.[0]?.delta?.reasoning_content ??
 				extractReasoningDetailsText(
-					data.choices?.[0]?.delta?.reasoning_details,
+					chunk.choices?.[0]?.delta?.reasoning_details,
 				) ??
 				""
 			);
+		}
 	}
 }

apps/gateway/src/chat/tools/extract-tool-calls.ts

@@ -11,6 +11,7 @@ export function extractToolCalls(
 ): any[] | null {
 	switch (provider) {
 		case "anthropic":
+		case "vertex-anthropic":
 			// Anthropic streaming tool calls come as content_block_start with tool_use type
 			if (
 				data.type === "content_block_start" &&

apps/gateway/src/chat/tools/parse-provider-response.ts

@@ -127,7 +127,8 @@ export function parseProviderResponse(
 
 			break;
 		}
-		case "anthropic": {
+		case "anthropic":
+		case "vertex-anthropic": {
 			// Extract content and reasoning content from Anthropic response
 			const contentBlocks = json.content ?? [];
 			const textBlocks = contentBlocks.filter(

apps/gateway/src/chat/tools/resolve-provider-context.ts

@@ -333,7 +333,8 @@ export async function resolveProviderContext(
 		usedProvider === "google-ai-studio" ||
 			usedProvider === "glacier" ||
 			usedProvider === "google-vertex" ||
-			usedProvider === "quartz"
+			usedProvider === "quartz" ||
+			usedProvider === "vertex-anthropic"
 			? usedToken
 			: undefined,
 		options.stream,
@@ -390,7 +391,7 @@ export async function resolveProviderContext(
 	}
 
 	// Anthropic does not allow temperature and top_p simultaneously
-	if (usedProvider === "anthropic") {
+	if (usedProvider === "anthropic" || usedProvider === "vertex-anthropic") {
 		if (temperature !== undefined && top_p !== undefined) {
 			top_p = undefined;
 		}

apps/gateway/src/chat/tools/transform-streaming-to-openai.ts

@@ -65,7 +65,8 @@ export function transformStreamingToOpenai(
 	};
 
 	switch (usedProvider) {
-		case "anthropic": {
+		case "anthropic":
+		case "vertex-anthropic": {
 			if (data.type === "content_block_delta" && data.delta?.text) {
 				transformedData = {
 					id: data.id ?? `chatcmpl-${Date.now()}`,

packages/actions/src/get-provider-endpoint.ts

@@ -162,6 +162,24 @@ export function getProviderEndpoint(
 						"https://aiplatform.googleapis.com",
 					) ?? "https://aiplatform.googleapis.com";
 				break;
+			case "vertex-anthropic": {
+				const vaDefaultRegion =
+					providerKeyOptions?.vertex_anthropic_region ??
+					getProviderEnvValue(
+						"vertex-anthropic",
+						"region",
+						configIndex,
+						"us-east5",
+					) ??
+					"us-east5";
+				url =
+					envValueOrDefault(
+						"vertex-anthropic",
+						"baseUrl",
+						`https://${vaDefaultRegion}-aiplatform.googleapis.com`,
+					) ?? `https://${vaDefaultRegion}-aiplatform.googleapis.com`;
+				break;
+			}
 			case "quartz":
 				url = skipEnvVars
 					? undefined
@@ -338,6 +356,30 @@ export function getProviderEndpoint(
 				configIndex,
 				providerKeyOptions,
 			);
+		case "vertex-anthropic": {
+			const vaProjectId =
+				providerKeyOptions?.vertex_anthropic_project_id ??
+				getProviderEnvValue("vertex-anthropic", "project", configIndex);
+			const vaRegion =
+				providerKeyOptions?.vertex_anthropic_region ??
+				getProviderEnvValue(
+					"vertex-anthropic",
+					"region",
+					configIndex,
+					"us-east5",
+				) ??
+				"us-east5";
+
+			if (!vaProjectId) {
+				throw new Error(
+					"LLM_VERTEX_ANTHROPIC_PROJECT environment variable is required for vertex-anthropic provider",
+				);
+			}
+
+			const vaModel = modelName ?? "claude-sonnet-4-6";
+			const vaEndpoint = stream ? "streamRawPredict" : "rawPredict";
+			return `${url}/v1/projects/${vaProjectId}/locations/${vaRegion}/publishers/anthropic/models/${vaModel}:${vaEndpoint}`;
+		}
 		case "perplexity":
 			return `${url}/chat/completions`;
 		case "novita":

packages/actions/src/get-provider-headers.ts

@@ -39,6 +39,11 @@ export function getProviderHeaders(
 		case "google-vertex":
 		case "quartz":
 			return requestIdHeader;
+		case "vertex-anthropic":
+			return {
+				...requestIdHeader,
+				Authorization: `Bearer ${token}`,
+			};
 		case "avalanche":
 			return {
 				...requestIdHeader,

packages/actions/src/prepare-request-body.ts

@@ -928,7 +928,9 @@ export async function prepareRequestBody(
 	// otherwise pass an unknown field through to OpenAI/Google/etc., risking a
 	// 400 from strict providers and confusing logs from lenient ones.
 	const providerHandlesCacheControl =
-		usedProvider === "anthropic" || usedProvider === "aws-bedrock";
+		usedProvider === "anthropic" ||
+		usedProvider === "vertex-anthropic" ||
+		usedProvider === "aws-bedrock";
 	if (!providerHandlesCacheControl) {
 		processedMessages = processedMessages.map((m) => {
 			if (!Array.isArray(m.content)) {
@@ -1322,7 +1324,8 @@ export async function prepareRequestBody(
 			}
 			break;
 		}
-		case "anthropic": {
+		case "anthropic":
+		case "vertex-anthropic": {
 			// Remove generic tool_choice that was added earlier
 			delete requestBody.tool_choice;
 
@@ -1611,6 +1614,11 @@ export async function prepareRequestBody(
 					// Note: The model capability (jsonOutput) should ensure the prompt guides JSON output.
 				}
 			}
+
+			if (usedProvider === "vertex-anthropic") {
+				requestBody.anthropic_version = "vertex-2023-10-16";
+				delete requestBody.model;
+			}
 			break;
 		}
 		case "aws-bedrock": {

packages/db/src/schema.ts

@@ -483,6 +483,8 @@ export interface ProviderKeyOptions {
 	azure_ai_foundry_api_version?: string;
 	alibaba_region?: "singapore" | "us-virginia" | "cn-beijing";
 	google_vertex_project_id?: string;
+	vertex_anthropic_project_id?: string;
+	vertex_anthropic_region?: string;
 }
 
 export const providerKey = pgTable(