Skip to content

🪢 Open source agent evals & observability: Trace, evaluate, and improve LLM applications with one open platform.

feat(in-app-agent): Foundation to make in-app agent available in OSS (self-hosted setups) (#16284)

Lysander Kiesele0c78264afa8
diff --git a/.env.dev.example b/.env.dev.example
index 7a377d2916..0381ecf79c 100644
--- a/.env.dev.example
+++ b/.env.dev.example
@@ -201,6 +201,9 @@ LANGFUSE_AWS_BEDROCK_REGION="eu-west-1"
LANGFUSE_AWS_BEDROCK_MODEL="eu.anthropic.claude-opus-4-8"
LANGFUSE_AWS_BEDROCK_SMALL_MODEL="eu.anthropic.claude-haiku-4-5-20251001-v1:0"
LANGFUSE_IN_APP_AGENT_AWS_PROFILE="playground"
+# Self-hosted Assistant instance switch. Unset or "false" hides Assistant.
+# Cloud is on unless this is "false".
+# LANGFUSE_IN_APP_AGENT_ENABLED="true"
# In-app agent sandbox. Local development defaults to the dangerous docker provider.
# Set to enable sandboxing locally. Leave unset to disable sandbox tools entirely.
diff --git a/docker-compose.build.yml b/docker-compose.build.yml
index 9f6439f9d1..822f82b65f 100644
--- a/docker-compose.build.yml
+++ b/docker-compose.build.yml
@@ -54,6 +54,19 @@ services:
REDIS_TLS_CA: ${REDIS_TLS_CA:-/certs/ca.crt}
REDIS_TLS_CERT: ${REDIS_TLS_CERT:-/certs/redis.crt}
REDIS_TLS_KEY: ${REDIS_TLS_KEY:-/certs/redis.key}
+ # Shared AI features configuration. These apply to AI features that run
+ # in both web and worker.
+ LANGFUSE_AWS_BEDROCK_MODEL: ${LANGFUSE_AWS_BEDROCK_MODEL:-}
+ LANGFUSE_AWS_BEDROCK_SMALL_MODEL: ${LANGFUSE_AWS_BEDROCK_SMALL_MODEL:-}
+ LANGFUSE_AWS_BEDROCK_REGION: ${LANGFUSE_AWS_BEDROCK_REGION:-}
+ LANGFUSE_IN_APP_AGENT_ENABLED: ${LANGFUSE_IN_APP_AGENT_ENABLED:-}
+ AWS_ACCESS_KEY_ID: ${AWS_ACCESS_KEY_ID:-}
+ AWS_SECRET_ACCESS_KEY: ${AWS_SECRET_ACCESS_KEY:-}
+ AWS_SESSION_TOKEN: ${AWS_SESSION_TOKEN:-}
+ AWS_PROFILE: ${AWS_PROFILE:-}
+ # In-app agent settings.
+ LANGFUSE_IN_APP_AGENT_MAX_ACTIVE_RUNS_PER_USER: ${LANGFUSE_IN_APP_AGENT_MAX_ACTIVE_RUNS_PER_USER:-}
+ LANGFUSE_IN_APP_AGENT_MAX_ACTIVE_RUNS_PER_ORG: ${LANGFUSE_IN_APP_AGENT_MAX_ACTIVE_RUNS_PER_ORG:-}
restart: always
healthcheck:
test: ["CMD", "wget", "--no-verbose", "--tries=1", "--spider", "http://127.0.0.1:3000/api/public/health"]
@@ -76,6 +89,17 @@ services:
- "3030:3030"
environment:
<<: *langfuse-web-env
+ # In-app agent worker settings (ignored by the web service).
+ # Compose defaults the queue consumer on; the worker schema defaults it
+ # off so ingestion-only processes do not consume agent jobs.
+ QUEUE_CONSUMER_IN_APP_AGENT_RUN_QUEUE_IS_ENABLED: ${QUEUE_CONSUMER_IN_APP_AGENT_RUN_QUEUE_IS_ENABLED:-true}
+ LANGFUSE_IN_APP_AGENT_RUN_QUEUE_PROCESSING_CONCURRENCY: ${LANGFUSE_IN_APP_AGENT_RUN_QUEUE_PROCESSING_CONCURRENCY:-}
+ LANGFUSE_IN_APP_AGENT_AWS_PROFILE: ${LANGFUSE_IN_APP_AGENT_AWS_PROFILE:-}
+ LANGFUSE_IN_APP_AGENT_SANDBOX_PROVIDER: ${LANGFUSE_IN_APP_AGENT_SANDBOX_PROVIDER:-}
+ LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_IMAGE_IDENTIFIER: ${LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_IMAGE_IDENTIFIER:-}
+ LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_EXECUTION_ROLE_ARN: ${LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_EXECUTION_ROLE_ARN:-}
+ LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_EGRESS_NETWORK_CONNECTOR_ARN: ${LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_EGRESS_NETWORK_CONNECTOR_ARN:-}
+ LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_REGION: ${LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_REGION:-}
restart: always
healthcheck:
test: ["CMD", "wget", "--no-verbose", "--tries=1", "--spider", "http://127.0.0.1:3030/api/health"]
diff --git a/docker-compose.yml b/docker-compose.yml
index 028eddf772..8657364684 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -73,6 +73,27 @@ services:
REDIS_TLS_KEY: ${REDIS_TLS_KEY:-/certs/redis.key}
EMAIL_FROM_ADDRESS: ${EMAIL_FROM_ADDRESS:-}
SMTP_CONNECTION_URL: ${SMTP_CONNECTION_URL:-}
+ # Shared AI features configuration. These apply to AI features that run
+ # in both web and worker.
+ LANGFUSE_IN_APP_AGENT_ENABLED: ${LANGFUSE_IN_APP_AGENT_ENABLED:-}
+ LANGFUSE_AWS_BEDROCK_MODEL: ${LANGFUSE_AWS_BEDROCK_MODEL:-}
+ LANGFUSE_AWS_BEDROCK_SMALL_MODEL: ${LANGFUSE_AWS_BEDROCK_SMALL_MODEL:-}
+ LANGFUSE_AWS_BEDROCK_REGION: ${LANGFUSE_AWS_BEDROCK_REGION:-}
+ AWS_ACCESS_KEY_ID: ${AWS_ACCESS_KEY_ID:-}
+ AWS_SECRET_ACCESS_KEY: ${AWS_SECRET_ACCESS_KEY:-}
+ AWS_SESSION_TOKEN: ${AWS_SESSION_TOKEN:-}
+ AWS_PROFILE: ${AWS_PROFILE:-}
+ # In-app agent worker settings (ignored by the web service).
+ # Compose defaults the queue consumer on; the worker schema defaults it
+ # off so ingestion-only processes do not consume agent jobs.
+ QUEUE_CONSUMER_IN_APP_AGENT_RUN_QUEUE_IS_ENABLED: ${QUEUE_CONSUMER_IN_APP_AGENT_RUN_QUEUE_IS_ENABLED:-true}
+ LANGFUSE_IN_APP_AGENT_RUN_QUEUE_PROCESSING_CONCURRENCY: ${LANGFUSE_IN_APP_AGENT_RUN_QUEUE_PROCESSING_CONCURRENCY:-}
+ LANGFUSE_IN_APP_AGENT_AWS_PROFILE: ${LANGFUSE_IN_APP_AGENT_AWS_PROFILE:-}
+ LANGFUSE_IN_APP_AGENT_SANDBOX_PROVIDER: ${LANGFUSE_IN_APP_AGENT_SANDBOX_PROVIDER:-}
+ LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_IMAGE_IDENTIFIER: ${LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_IMAGE_IDENTIFIER:-}
+ LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_EXECUTION_ROLE_ARN: ${LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_EXECUTION_ROLE_ARN:-}
+ LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_EGRESS_NETWORK_CONNECTOR_ARN: ${LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_EGRESS_NETWORK_CONNECTOR_ARN:-}
+ LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_REGION: ${LANGFUSE_IN_APP_AGENT_SANDBOX_AWS_LAMBDA_MICROVM_REGION:-}
langfuse-web:
image: docker.langfuse.com/langfuse/langfuse:4
@@ -93,6 +114,9 @@ services:
LANGFUSE_INIT_USER_EMAIL: ${LANGFUSE_INIT_USER_EMAIL:-}
LANGFUSE_INIT_USER_NAME: ${LANGFUSE_INIT_USER_NAME:-}
LANGFUSE_INIT_USER_PASSWORD: ${LANGFUSE_INIT_USER_PASSWORD:-}
+ # In-app agent settings.
+ LANGFUSE_IN_APP_AGENT_MAX_ACTIVE_RUNS_PER_USER: ${LANGFUSE_IN_APP_AGENT_MAX_ACTIVE_RUNS_PER_USER:-}
+ LANGFUSE_IN_APP_AGENT_MAX_ACTIVE_RUNS_PER_ORG: ${LANGFUSE_IN_APP_AGENT_MAX_ACTIVE_RUNS_PER_ORG:-}
clickhouse:
image: docker.io/clickhouse/clickhouse-server:25.12
diff --git a/packages/shared/AGENTS.md b/packages/shared/AGENTS.md
index 0f7b5ca3a8..7791acd883 100644
--- a/packages/shared/AGENTS.md
+++ b/packages/shared/AGENTS.md
@@ -45,7 +45,8 @@
for shared backend services, repositories, queue helpers/contracts, Redis and
ClickHouse helpers, auth helpers, logger/instrumentation, ingestion helpers,
AI SDK-native LLM execution helpers (`generateLLMText` and
- `streamLLMText`), and server test utilities.
+ `streamLLMText`), Bedrock default-credential provider auth
+ (`createDefaultBedrockProviderAuth`), and server test utilities.
- `@langfuse/shared/src/db` via `src/db.ts`: Prisma client singleton plus
Prisma namespace/types for direct database access. Never route this into
frontend-safe code.
@@ -60,12 +61,13 @@
and interrupt parsing. Never re-export server code here.
- In-app-agent server contracts use explicit subpaths only:
`persistence`, `runLifecycle`, `tunables`, `eventCompaction`, `mcpPolicy`,
- `toolResults`, `toolErrors`, and `systemPrompt`. These are storage/lifecycle or durable
- cross-process policy contracts; the Mastra runtime and sandbox belong to the
- worker.
+ `toolResults`, `toolErrors`, `systemPrompt`, and `modelProvider`. These are
+ storage/lifecycle, durable cross-process policy, or instance-model contracts;
+ the Mastra runtime and sandbox belong to the worker.
- Narrower exported subpaths also exist for targeted imports:
`@langfuse/shared/src/server/auth/apiKeys`,
- `@langfuse/shared/src/server/ee/ingestionMasking`, and
+ `@langfuse/shared/src/server/ee/ingestionMasking`,
+ `@langfuse/shared/src/server/llm/llmText`, and
`@langfuse/shared/src/utils/chatml`.
When changing export surfaces, keep `package.json#exports`, the relevant barrel
diff --git a/packages/shared/package.json b/packages/shared/package.json
index 91538fab2e..7d0fd39cce 100644
--- a/packages/shared/package.json
+++ b/packages/shared/package.json
@@ -31,6 +31,10 @@
"import": "./dist/src/server/index.js",
"require": "./dist/src/server/index.js"
},
+ "./src/server/llm/llmText": {
+ "import": "./dist/src/server/llm/llmText.js",
+ "require": "./dist/src/server/llm/llmText.js"
+ },
"./src/server/auth/apiKeys": {
"import": "./dist/src/server/auth/apiKeys.js",
"require": "./dist/src/server/auth/apiKeys.js"
@@ -98,6 +102,10 @@
"./in-app-agent/server/systemPrompt": {
"import": "./dist/src/in-app-agent/server/systemPrompt.js",
"require": "./dist/src/in-app-agent/server/systemPrompt.js"
+ },
+ "./in-app-agent/server/modelProvider": {
+ "import": "./dist/src/in-app-agent/server/modelProvider.js",
+ "require": "./dist/src/in-app-agent/server/modelProvider.js"
}
},
"scripts": {
diff --git a/packages/shared/src/env.ts b/packages/shared/src/env.ts
index 8fd573c109..53ec693f3d 100644
--- a/packages/shared/src/env.ts
+++ b/packages/shared/src/env.ts
@@ -522,6 +522,7 @@ const EnvSchema = z.object({
LANGFUSE_AWS_BEDROCK_REGION: z.string().optional(),
LANGFUSE_AWS_BEDROCK_MODEL: z.string().optional(),
LANGFUSE_AWS_BEDROCK_SMALL_MODEL: z.string().optional(),
+ LANGFUSE_IN_APP_AGENT_ENABLED: z.enum(["true", "false"]).optional(),
// API Performance Flags
// Whether to add a `FINAL` modifier to the observations CTE in GET /api/public/traces.
diff --git a/packages/shared/src/in-app-agent/server/modelProvider.ts b/packages/shared/src/in-app-agent/server/modelProvider.ts
new file mode 100644
index 0000000000..557a12f0ec
--- /dev/null
+++ b/packages/shared/src/in-app-agent/server/modelProvider.ts
@@ -0,0 +1,54 @@
+import { env } from "../../env";
+import { assertValidBedrockRegion } from "../../server/llm/ai-sdk/providers/bedrock";
+
+export type InAppAgentModelConfig = {
+ provider: "bedrock";
+ modelId: string;
+ titleModelId: string;
+ region: string;
+};
+
+/**
+ * Resolves the instance-wide in-app agent model configuration.
+ *
+ * The discriminated provider contract is intentionally narrower than the
+ * generic LLM connection model: in-app agent execution currently supports
+ * Bedrock only, but callers do not need to know which provider supplies the
+ * model. A future provider can add another branch here without changing the
+ * worker or title-generation call sites.
+ */
+export function getInAppAgentModelConfig(params?: {
+ modelId?: string | null;
+}): InAppAgentModelConfig | undefined {
+ const modelId = params?.modelId ?? env.LANGFUSE_AWS_BEDROCK_MODEL;
+ const region = env.LANGFUSE_AWS_BEDROCK_REGION;
+
+ if (!modelId || !region) {
+ return undefined;
+ }
+
+ assertValidBedrockRegion(region);
+
+ return {
+ provider: "bedrock",
+ modelId,
+ titleModelId: env.LANGFUSE_AWS_BEDROCK_SMALL_MODEL ?? modelId,
+ region,
+ };
+}
+
+/**
+ * Instance-wide in-app agent switch.
+ *
+ * Cloud is on unless LANGFUSE_IN_APP_AGENT_ENABLED is "false". Self-hosted is
+ * on only when the var is "true".
+ */
+export function isInAppAgentInstanceEnabled(): boolean {
+ const enabled = env.LANGFUSE_IN_APP_AGENT_ENABLED;
+
+ if (env.NEXT_PUBLIC_LANGFUSE_CLOUD_REGION) {
+ return enabled !== "false";
+ }
+
+ return enabled === "true";
+}
diff --git a/packages/shared/src/in-app-agent/server/persistence.ts b/packages/shared/src/in-app-agent/server/persistence.ts
index b6ee67cf6a..a2b9646e87 100644
--- a/packages/shared/src/in-app-agent/server/persistence.ts
+++ b/packages/shared/src/in-app-agent/server/persistence.ts
@@ -15,11 +15,11 @@ import {
import { Prisma } from "../../db";
import type { InAppAgentConversation, PrismaClient } from "../../db";
-import { env } from "../../env";
import {
generateLangfuseAIText,
getLangfuseAITraceSinkParams,
} from "../../server/llm/langfuseAiCompletion";
+import { getInAppAgentModelConfig } from "./modelProvider";
import { getProductBaseUrl } from "../../server/utils/baseUrl";
import { truncate } from "../../utils/stringChecks";
import { assertUnreachable } from "../../utils/typeChecks";
@@ -431,10 +431,9 @@ export async function maybeInferAndPersistConversationTitle(params: {
userId: string;
aiTelemetryEnabled: boolean;
}) {
- const model =
- env.LANGFUSE_AWS_BEDROCK_SMALL_MODEL ?? env.LANGFUSE_AWS_BEDROCK_MODEL;
+ const modelConfig = getInAppAgentModelConfig();
- if (!model) {
+ if (!modelConfig) {
return;
}
@@ -523,7 +522,7 @@ ${JSON.stringify(transcript, null, 2)}
`.trim(),
},
],
- model,
+ model: modelConfig.titleModelId,
maxTokens: 1000,
traceSinkParams: params.aiTelemetryEnabled
? getLangfuseAITraceSinkParams({
diff --git a/packages/shared/src/server/index.ts b/packages/shared/src/server/index.ts
index b3308bb988..7ec6aacff3 100644
--- a/packages/shared/src/server/index.ts
+++ b/packages/shared/src/server/index.ts
@@ -57,6 +57,7 @@ export * from "./utils/productUrl";
export * from "./llm/compileChatMessages";
export * from "./llm/testModelCall";
export * from "./llm/baseUrlValidation";
+export * from "./llm/ai-sdk/providers/bedrock";
export * from "./outbound-url";
export * from "./services/blobStorageEndpointValidation";
export * from "./llm/getInternalTracingHandler";
diff --git a/packages/shared/src/server/llm/ai-sdk/providers/bedrock.ts b/packages/shared/src/server/llm/ai-sdk/providers/bedrock.ts
index 84f234b669..5940830a56 100644
--- a/packages/shared/src/server/llm/ai-sdk/providers/bedrock.ts
+++ b/packages/shared/src/server/llm/ai-sdk/providers/bedrock.ts
@@ -80,6 +80,28 @@ export function translateBedrockProviderOptions(
// requestShape.test.ts.
const SUPPRESS_BEARER_TOKEN_ENV_FALLBACK = { apiKey: "" };
+/**
+ * Returns the Bedrock auth options for the AWS default credential chain.
+ *
+ * The empty API key deliberately prevents the AI SDK from selecting an
+ * ambient AWS_BEARER_TOKEN_BEDROCK value ahead of the configured AWS
+ * credentials. An optional profile is useful for local worker development;
+ * production workloads should normally rely on the default AWS role chain.
+ */
+export function createDefaultBedrockProviderAuth(params?: {
+ profile?: string;
+}): Pick<
+ Parameters<typeof createAmazonBedrock>[0] & object,
+ "apiKey" | "credentialProvider"
+> {
+ return {
+ credentialProvider: fromNodeProviderChain(
+ params?.profile ? { profile: params.profile } : {},
+ ),
+ ...SUPPRESS_BEARER_TOKEN_ENV_FALLBACK,
+ };
+}
+
/**
* Resolves the persisted Bedrock credential contract: the decrypted secret is
* either the default-credentials sentinel (allowed only self-hosted or for
@@ -101,10 +123,7 @@ export function resolveBedrockProviderAuth(params: {
) {
// Unlike the AI SDK's built-in env-only fallback, the node provider chain
// includes env, profile, IMDS, IRSA, and the remaining AWS defaults.
- return {
- credentialProvider: fromNodeProviderChain(),
- ...SUPPRESS_BEARER_TOKEN_ENV_FALLBACK,
- };
+ return createDefaultBedrockProviderAuth();
}
try {
diff --git a/pnpm-lock.yaml b/pnpm-lock.yaml
index 257829c883..b2e2fb3f12 100644
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -1021,9 +1021,6 @@ importers:
'@aws-sdk/client-lambda-microvms':
specifier: 'catalog:'
version: 3.1074.0
- '@aws-sdk/credential-providers':
- specifier: 'catalog:'
- version: 3.1074.0
'@langfuse/shared':
specifier: workspace:*
version: link:../packages/shared
diff --git a/web/.structure-baseline.json b/web/.structure-baseline.json
index de4afebda5..25fbf4973e 100644
--- a/web/.structure-baseline.json
+++ b/web/.structure-baseline.json
@@ -233,7 +233,7 @@
"src/features/evals/components/evaluation-prompt-preview.tsx: extra exports: getVariableColor",
"src/features/experiments/components/table/ExperimentItemsTable.tsx: extra exports: getDefaultExperimentFilterTarget, shouldEnableExperimentPeek",
"src/features/in-app-agent/components/InAppAgentWindowShell.tsx: extra exports: useInAppAgentWindowShellPanelControl",
- "src/features/in-app-agent/components/InAppAiAgentProvider.tsx: extra exports: useInAppAiAgent, useCanUseInAppAgent",
+ "src/features/in-app-agent/components/InAppAiAgentProvider.tsx: extra exports: useInAppAiAgent, useCanUseInAppAgent, useIsInAppAgentLauncherVisible",
"src/features/monitors/components/MonitorChartPreview.tsx: extra exports: __test",
"src/features/monitors/components/MonitorForm.tsx: extra exports: __test",
"src/features/monitors/components/MonitorsTable.tsx: extra exports: __test",
diff --git a/web/src/__tests__/server/in-app-agent-background-run.servertest.ts b/web/src/__tests__/server/in-app-agent-background-run.servertest.ts
index d7cc088b89..dcecd86182 100644
--- a/web/src/__tests__/server/in-app-agent-background-run.servertest.ts
+++ b/web/src/__tests__/server/in-app-agent-background-run.servertest.ts
@@ -12,6 +12,7 @@ import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
import { type Plan } from "@langfuse/shared";
import { prisma } from "@langfuse/shared/src/db";
+import { env as sharedEnv } from "@langfuse/shared/src/env";
import { createOrgProjectAndApiKey } from "@langfuse/shared/src/server";
import {
IN_APP_AGENT_APPROVAL_DECISION_EVENT_NAME,
@@ -96,6 +97,11 @@ vi.mock("@/src/server/auth", () => ({
describe("in-app agent background runs", () => {
const originalCloudRegion = env.NEXT_PUBLIC_LANGFUSE_CLOUD_REGION;
const originalBedrockModel = env.LANGFUSE_AWS_BEDROCK_MODEL;
+ const originalSharedCloudRegion = sharedEnv.NEXT_PUBLIC_LANGFUSE_CLOUD_REGION;
+ const originalSharedBedrockModel = sharedEnv.LANGFUSE_AWS_BEDROCK_MODEL;
+ const originalSharedBedrockRegion = sharedEnv.LANGFUSE_AWS_BEDROCK_REGION;
+ const originalSharedInAppAgentEnabled =
+ sharedEnv.LANGFUSE_IN_APP_AGENT_ENABLED;
const originalMaxActiveRunsPerUser =
env.LANGFUSE_IN_APP_AGENT_MAX_ACTIVE_RUNS_PER_USER;
const originalMaxActiveRunsPerOrg =
@@ -104,6 +110,10 @@ describe("in-app agent background runs", () => {
beforeEach(() => {
(env as any).NEXT_PUBLIC_LANGFUSE_CLOUD_REGION = "DEV";
(env as any).LANGFUSE_AWS_BEDROCK_MODEL = "test-model";
+ (sharedEnv as any).NEXT_PUBLIC_LANGFUSE_CLOUD_REGION = "DEV";
+ (sharedEnv as any).LANGFUSE_AWS_BEDROCK_MODEL = "test-model";
+ (sharedEnv as any).LANGFUSE_AWS_BEDROCK_REGION = "eu-central-1";
+ (sharedEnv as any).LANGFUSE_IN_APP_AGENT_ENABLED = undefined;
enqueuedJobs.length = 0;
enqueueShouldFail = false;
persistenceMocks.maybeInferAndPersistConversationTitle.mockClear();
@@ -117,6 +127,13 @@ describe("in-app agent background runs", () => {
vi.useRealTimers();
(env as any).NEXT_PUBLIC_LANGFUSE_CLOUD_REGION = originalCloudRegion;
(env as any).LANGFUSE_AWS_BEDROCK_MODEL = originalBedrockModel;
+ (sharedEnv as any).NEXT_PUBLIC_LANGFUSE_CLOUD_REGION =
+ originalSharedCloudRegion;
+ (sharedEnv as any).LANGFUSE_AWS_BEDROCK_MODEL = originalSharedBedrockModel;
+ (sharedEnv as any).LANGFUSE_AWS_BEDROCK_REGION =
+ originalSharedBedrockRegion;
+ (sharedEnv as any).LANGFUSE_IN_APP_AGENT_ENABLED =
+ originalSharedInAppAgentEnabled;
(env as any).LANGFUSE_IN_APP_AGENT_MAX_ACTIVE_RUNS_PER_USER =
originalMaxActiveRunsPerUser;
(env as any).LANGFUSE_IN_APP_AGENT_MAX_ACTIVE_RUNS_PER_ORG =
@@ -194,6 +211,103 @@ describe("in-app agent background runs", () => {
userId: params.userId,
});
+ it.each([
+ {
+ name: "starts a self-hosted run when LANGFUSE_IN_APP_AGENT_ENABLED is true",
+ cloudRegion: undefined,
+ enabled: "true" as const,
+ plan: "self-hosted:enterprise" as const,
+ expected: "queued" as const,
+ },
+ {
+ name: "rejects a self-hosted run when LANGFUSE_IN_APP_AGENT_ENABLED is unset",
+ cloudRegion: undefined,
+ enabled: undefined,
+ plan: "self-hosted:enterprise" as const,
+ expected: "rejected" as const,
+ },
+ {
+ name: "rejects a self-hosted run when LANGFUSE_IN_APP_AGENT_ENABLED is false",
+ cloudRegion: undefined,
+ enabled: "false" as const,
+ plan: "self-hosted:enterprise" as const,
+ expected: "rejected" as const,
+ },
+ {
+ name: "starts a Cloud run when LANGFUSE_IN_APP_AGENT_ENABLED is unset",
+ cloudRegion: "DEV" as const,
+ enabled: undefined,
+ plan: "cloud:hobby" as const,
+ expected: "queued" as const,
+ },
+ {
+ name: "rejects a Cloud run when LANGFUSE_IN_APP_AGENT_ENABLED is false",
+ cloudRegion: "DEV" as const,
+ enabled: "false" as const,
+ plan: "cloud:hobby" as const,
+ expected: "rejected" as const,
+ },
+ ])("$name", async ({ cloudRegion, enabled, plan, expected }) => {
+ (env as any).NEXT_PUBLIC_LANGFUSE_CLOUD_REGION = cloudRegion;
+ (sharedEnv as any).NEXT_PUBLIC_LANGFUSE_CLOUD_REGION = cloudRegion;
+ (sharedEnv as any).LANGFUSE_IN_APP_AGENT_ENABLED = enabled;
+ const { caller, projectId, userId } = await createCaller(undefined, plan);
+ const conversation = await createConversation({ projectId, userId });
+
+ const startRun = caller.startRun({
+ projectId,
+ conversationId: conversation.id,
+ message: "Inspect a trace",
+ });
+
+ if (expected === "queued") {
+ await expect(startRun).resolves.toMatchObject({
+ conversationId: conversation.id,
+ });
+ expect(enqueuedJobs).toHaveLength(1);
+ return;
+ }
+
+ await expect(startRun).rejects.toMatchObject({
+ code: "PRECONDITION_FAILED",
+ message: "In-app agent is not enabled on this instance.",
+ });
+ expect(enqueuedJobs).toHaveLength(0);
+ });
+
+ it("lists conversations when the in-app agent model is not configured", async () => {
+ const { caller, projectId, userId } = await createCaller();
+ const conversation = await createConversation({ projectId, userId });
+ (sharedEnv as any).LANGFUSE_AWS_BEDROCK_MODEL = undefined;
+ (sharedEnv as any).LANGFUSE_AWS_BEDROCK_REGION = undefined;
+
+ const listed = await caller.listConversations({ projectId, limit: 50 });
+
+ expect(listed.conversations.map((item) => item.id)).toEqual([
+ conversation.id,
+ ]);
+ });
+
+ it("rejects requests before queueing when the in-app agent model is not configured", async () => {
+ const { caller, projectId } = await createCaller();
+ (sharedEnv as any).LANGFUSE_AWS_BEDROCK_MODEL = undefined;
+ (sharedEnv as any).LANGFUSE_AWS_BEDROCK_REGION = undefined;
+
+ await expect(
+ caller.startRun({
+ projectId,
+ conversationId: createInAppAgentConversationId(),
+ message: "Do not queue this",
+ }),
+ ).rejects.toMatchObject({
+ code: "PRECONDITION_FAILED",
+ message:
+ "In-app agent Bedrock model is not configured. Set LANGFUSE_AWS_BEDROCK_MODEL and LANGFUSE_AWS_BEDROCK_REGION.",
+ });
+
+ expect(enqueuedJobs).toHaveLength(0);
+ });
+
/** Park a run for approval the way the worker does: interrupt event, then CAS. */
const parkRunForApproval = async (params: {
projectId: string;
diff --git a/web/src/__tests__/server/in-app-agent-persistence.servertest.ts b/web/src/__tests__/server/in-app-agent-persistence.servertest.ts
index c74171bac6..a2345f9693 100644
--- a/web/src/__tests__/server/in-app-agent-persistence.servertest.ts
+++ b/web/src/__tests__/server/in-app-agent-persistence.servertest.ts
@@ -1,5 +1,7 @@
-vi.mock("@langfuse/shared/src/server", async () => {
- const actual = await vi.importActual("@langfuse/shared/src/server");
+vi.mock("@langfuse/shared/src/server/llm/llmText", async () => {
+ const actual = await vi.importActual(
+ "@langfuse/shared/src/server/llm/llmText",
+ );
return {
...actual,
generateLLMText: vi.fn(),
@@ -17,9 +19,9 @@ import { type Plan } from "@langfuse/shared";
import { prisma } from "@langfuse/shared/src/db";
import {
createOrgProjectAndApiKey,
- generateLLMText,
getScoreById,
} from "@langfuse/shared/src/server";
+import { generateLLMText } from "@langfuse/shared/src/server/llm/llmText";
import { env } from "@/src/env.mjs";
import {
InAppAgentRunErrorCode,
@@ -244,18 +246,16 @@ describe("in-app agent persistence", () => {
});
});
- it("rejects users without the in-app agent entitlement", async () => {
+ it("allows oss-plan users to list conversations", async () => {
const { caller, projectId } = await createCaller(
`user-${randomUUID()}`,
"oss",
);
- await expect(caller.listConversations({ projectId })).rejects.toMatchObject(
- {
- code: "FORBIDDEN",
- message: expect.stringContaining("in-app-agent"),
- },
- );
+ await expect(caller.listConversations({ projectId })).resolves.toEqual({
+ conversations: [],
+ nextCursor: undefined,
+ });
});
const startCompactRun = async (params: {
diff --git a/web/src/__tests__/server/organization-ai-features.servertest.ts b/web/src/__tests__/server/organization-ai-features.servertest.ts
new file mode 100644
index 0000000000..1f24c25c50
--- /dev/null
+++ b/web/src/__tests__/server/organization-ai-features.servertest.ts
@@ -0,0 +1,154 @@
+import { randomUUID } from "crypto";
+import type { Session } from "next-auth";
+import { afterEach, beforeEach, describe, expect, it } from "vitest";
+
+import { prisma } from "@langfuse/shared/src/db";
+import { env as sharedEnv } from "@langfuse/shared/src/env";
+import { env } from "@/src/env.mjs";
+import { appRouter } from "@/src/server/api/root";
+import { createInnerTRPCContext } from "@/src/server/api/trpc";
+
+describe("organization AI feature settings", () => {
+ const originalCloudRegion = env.NEXT_PUBLIC_LANGFUSE_CLOUD_REGION;
+ const originalSharedInAppAgentEnabled =
+ sharedEnv.LANGFUSE_IN_APP_AGENT_ENABLED;
+
+ beforeEach(() => {
+ (env as any).NEXT_PUBLIC_LANGFUSE_CLOUD_REGION = undefined;
+ });
+
+ afterEach(() => {
+ (env as any).NEXT_PUBLIC_LANGFUSE_CLOUD_REGION = originalCloudRegion;
+ (sharedEnv as any).LANGFUSE_IN_APP_AGENT_ENABLED =
+ originalSharedInAppAgentEnabled;
+ });
+
+ it("allows a self-hosted organization administrator to opt in to AI features", async () => {
+ const { caller, orgId } = await createCaller();
+
+ await caller.organizations.update({
+ orgId,
+ aiFeaturesEnabled: true,
+ });
+
+ await expect(
+ prisma.organization.findUniqueOrThrow({
+ where: { id: orgId },
+ select: { aiFeaturesEnabled: true },
+ }),
+ ).resolves.toEqual({ aiFeaturesEnabled: true });
+ });
+
+ it("allows a self-hosted organization administrator to opt in to AI features when in-app agent is instance-disabled", async () => {
+ (sharedEnv as any).LANGFUSE_IN_APP_AGENT_ENABLED = "false";
+ const { caller, orgId } = await createCaller();
+
+ await caller.organizations.update({
+ orgId,
+ aiFeaturesEnabled: true,
+ });
+
+ await expect(
+ prisma.organization.findUniqueOrThrow({
+ where: { id: orgId },
+ select: { aiFeaturesEnabled: true },
+ }),
+ ).resolves.toEqual({ aiFeaturesEnabled: true });
+ });
+
+ it("keeps AI telemetry controls Cloud-only", async () => {
+ const { caller, orgId } = await createCaller();
+
+ await expect(
+ caller.organizations.update({
+ orgId,
+ aiTelemetryEnabled: false,
+ }),
+ ).rejects.toMatchObject({
+ code: "PRECONDITION_FAILED",
+ message: "AI telemetry controls are only available on Langfuse Cloud.",
+ });
+ });
+});
+
+async function createCaller() {
+ const id = randomUUID();
+ const orgId = `org-${id}`;
+ const projectId = `project-${id}`;
+ const userId = `user-${id}`;
+
+ const organization = await prisma.organization.create({
+ data: {
+ id: orgId,
+ name: `Organization AI Features Test ${id}`,
+ },
+ });
+ const project = await prisma.project.create({
+ data: {
+ id: projectId,
+ orgId,
+ name: "Organization AI Features Test Project",
+ },
+ });
+ const user = await prisma.user.create({
+ data: {
+ id: userId,
+ email: `${userId}@example.com`,
+ name: "Organization AI Features Test User",
+ },
+ });
+
+ const session: Session = {
+ expires: "1",
+ user: {
+ id: user.id,
+ email: user.email,
+ name: user.name,
+ canCreateOrganizations: true,
+ organizations: [
+ {
+ id: organization.id,
+ name: organization.name,
+ role: "OWNER",
+ plan: "self-hosted:enterprise",
+ cloudConfig: undefined,
+ metadata: {},
+ aiFeaturesEnabled: false,
+ aiTelemetryEnabled: true,
+ projects: [
+ {
+ id: project.id,
+ name: project.name,
+ role: "ADMIN",
+ deletedAt: null,
+ retentionDays: null,
+ hasTraces: false,
+ metadata: {},
+ createdAt: project.createdAt.toISOString(),
+ },
+ ],
+ },
+ ],
+ featureFlags: {
+ searchBar: false,
+ templateFlag: false,
+ excludeClickhouseRead: false,
+ observationEvals: false,
+ v4BetaToggleVisible: false,
+ experimentsV4Enabled: false,
+ },
+ admin: false,
+ },
+ environment: {
+ enableExperimentalFeatures: false,
+ selfHostedInstancePlan: "self-hosted:enterprise",
+ },
+ };
+
+ const ctx = createInnerTRPCContext({ session, headers: {} });
+
+ return {
+ orgId,
+ caller: appRouter.createCaller({ ...ctx, prisma }),
+ };
+}
diff --git a/web/src/components/nav/in-app-ai-agent-button.clienttest.tsx b/web/src/components/nav/in-app-ai-agent-button.clienttest.tsx
index 10c23105e1..2c75d6b7da 100644
--- a/web/src/components/nav/in-app-ai-agent-button.clienttest.tsx
+++ b/web/src/components/nav/in-app-ai-agent-button.clienttest.tsx
@@ -10,7 +10,7 @@ const mocks = vi.hoisted(() => ({
}));
vi.mock("@/src/features/in-app-agent/components/InAppAiAgentProvider", () => ({
- useCanUseInAppAgent: () => true,
+ useIsInAppAgentLauncherVisible: () => true,
useInAppAiAgent: () => ({
open: mocks.open,
attentionCount: mocks.attentionCount,
diff --git a/web/src/components/nav/in-app-ai-agent-button.tsx b/web/src/components/nav/in-app-ai-agent-button.tsx
index bdf7ca677a..6ac1e2d737 100644
--- a/web/src/components/nav/in-app-ai-agent-button.tsx
+++ b/web/src/components/nav/in-app-ai-agent-button.tsx
@@ -4,7 +4,7 @@ import { BotMessageSquare } from "lucide-react";
import { Button } from "@/src/components/ui/button";
import { KeyboardShortcut } from "@/src/components/ui/keyboard-shortcut";
import {
- useCanUseInAppAgent,
+ useIsInAppAgentLauncherVisible,
useInAppAiAgent,
type InAppAgentEntryPoint,
} from "@/src/features/in-app-agent/components/InAppAiAgentProvider";
@@ -24,7 +24,7 @@ export const InAppAiAgentButton = ({
prominent?: boolean;
} = {}) => {
const { open, setOpen, openAssistant, attentionCount } = useInAppAiAgent();
- const canUseAssistant = useCanUseInAppAgent();
+ const isInAppAgentLauncherVisible = useIsInAppAgentLauncherVisible();
const toggleAssistant = useCallback(
(source: InAppAgentEntryPoint) => {
@@ -39,7 +39,7 @@ export const InAppAiAgentButton = ({
);
useEffect(() => {
- if (!canUseAssistant) {
+ if (!isInAppAgentLauncherVisible) {
return;
}
@@ -60,9 +60,9 @@ export const InAppAiAgentButton = ({
document.addEventListener("keydown", handleKeyDown);
return () => document.removeEventListener("keydown", handleKeyDown);
- }, [canUseAssistant, toggleAssistant]);
+ }, [isInAppAgentLauncherVisible, toggleAssistant]);
- if (!canUseAssistant) {
+ if (!isInAppAgentLauncherVisible) {
return null;
}
diff --git a/web/src/env.mjs b/web/src/env.mjs
index 05d95b2243..0d907139bf 100644
--- a/web/src/env.mjs
+++ b/web/src/env.mjs
@@ -452,6 +452,7 @@ export const env = createEnv({
// AWS Bedrock for langfuse native AI feature such as natural language filters
LANGFUSE_AWS_BEDROCK_MODEL: z.string().optional(),
LANGFUSE_AWS_BEDROCK_SMALL_MODEL: z.string().optional(),
+ LANGFUSE_IN_APP_AGENT_ENABLED: z.enum(["true", "false"]).optional(),
// Tracing for Langfuse AI Features
LANGFUSE_AI_FEATURES_HOST: z.string().optional(),
@@ -977,6 +978,7 @@ export const env = createEnv({
LANGFUSE_AWS_BEDROCK_MODEL: process.env.LANGFUSE_AWS_BEDROCK_MODEL,
LANGFUSE_AWS_BEDROCK_SMALL_MODEL:
process.env.LANGFUSE_AWS_BEDROCK_SMALL_MODEL,
+ LANGFUSE_IN_APP_AGENT_ENABLED: process.env.LANGFUSE_IN_APP_AGENT_ENABLED,
// Langfuse Tracing AI Features
LANGFUSE_AI_FEATURES_HOST: process.env.LANGFUSE_AI_FEATURES_HOST,
diff --git a/web/src/features/entitlements/constants/entitlements.ts b/web/src/features/entitlements/constants/entitlements.ts
index 1bac2d9c86..39e880d326 100644
--- a/web/src/features/entitlements/constants/entitlements.ts
+++ b/web/src/features/entitlements/constants/entitlements.ts
@@ -30,6 +30,7 @@ const cloudAllPlansEntitlements: Entitlement[] = [
const selfHostedAllPlansEntitlements: Entitlement[] = [
"trace-deletion",
"scheduled-blob-exports",
+ "in-app-agent",
];
// Entitlement Limits: Limits on the number of resources that can be created/used
diff --git a/web/src/features/evals/pages/remap-evaluator.tsx b/web/src/features/evals/pages/remap-evaluator.tsx
index 2fa3f844a2..629494a47a 100644
--- a/web/src/features/evals/pages/remap-evaluator.tsx
+++ b/web/src/features/evals/pages/remap-evaluator.tsx
@@ -22,7 +22,7 @@ import {
import { BotMessageSquare, ChevronDown, Zap } from "lucide-react";
import { useEvalCapabilities } from "@/src/features/evals/hooks/useEvalCapabilities";
import {
- useCanUseInAppAgent,
+ useIsInAppAgentLauncherVisible,
useInAppAiAgent,
} from "@/src/features/in-app-agent/components/InAppAiAgentProvider";
import { usePostHogClientCapture } from "@/src/features/posthog-analytics/usePostHogClientCapture";
@@ -48,7 +48,7 @@ export default function RemapEvaluatorPage() {
const evalConfigId = router.query.evaluator as string;
const v4UpgradeUiEnabled = useV4UpgradeUiEnabled(projectId);
const capture = usePostHogClientCapture();
- const canUseAssistant = useCanUseInAppAgent();
+ const isInAppAgentLauncherVisible = useIsInAppAgentLauncherVisible();
const { openAssistant, submit } = useInAppAiAgent();
const sdk = useProjectV4SdkData({
projectId,
@@ -231,7 +231,7 @@ export default function RemapEvaluatorPage() {
align="top"
actions={() => (
<>
- {canUseAssistant ? (
+ {isInAppAgentLauncherVisible ? (
<Button
size="sm"
variant="secondary"
diff --git a/web/src/features/in-app-agent/components/InAppAgentExecution.clienttest.tsx b/web/src/features/in-app-agent/components/InAppAgentExecution.clienttest.tsx
index 42d3f37f6a..6e55f721a1 100644
--- a/web/src/features/in-app-agent/components/InAppAgentExecution.clienttest.tsx
+++ b/web/src/features/in-app-agent/components/InAppAgentExecution.clienttest.tsx
@@ -112,6 +112,7 @@ vi.mock("next/router", () => ({
const sessionMocks = vi.hoisted(() => ({
userId: "user-1" as string | undefined,
aiFeaturesEnabled: true,
+ inAppAgentEnabled: true,
isLangfuseCloud: true,
}));
@@ -119,6 +120,7 @@ vi.mock("next-auth/react", () => ({
useSession: () => ({
data: {
user: { id: sessionMocks.userId, name: "Test User" },
+ environment: { inAppAgentEnabled: sessionMocks.inAppAgentEnabled },
},
}),
}));
@@ -279,6 +281,7 @@ describe("in-app agent execution", () => {
beforeEach(() => {
sessionMocks.userId = "user-1";
sessionMocks.aiFeaturesEnabled = true;
+ sessionMocks.inAppAgentEnabled = true;
sessionMocks.isLangfuseCloud = true;
providerMocks.activityUseQuery.mockImplementation(
() => providerMocks.activityQuery,
@@ -301,9 +304,10 @@ describe("in-app agent execution", () => {
window.sessionStorage.clear();
});
- // Gated more strictly than useCanUseInAppAgent, which ignores the org AI
- // toggle so the entry points can offer to turn it on. Polling with it off
- // turns every page load and window focus into a Forbidden toast.
+ // Polling uses useCanUseInAppAgent (org AI Features on). Entry points use
+ // useIsInAppAgentLauncherVisible, which still shows when AI Features are off
+ // so clicking can open the enable dialog. Polling with it off turns every
+ // page load and window focus into a Forbidden toast.
it("polls for activity only once AI features are on", () => {
sessionMocks.aiFeaturesEnabled = false;
@@ -330,6 +334,20 @@ describe("in-app agent execution", () => {
);
});
+ it("does not poll for activity when in-app agent is instance-disabled", () => {
+ sessionMocks.inAppAgentEnabled = false;
+
+ render(
+ <InAppAiAgentProvider defaultOpen={false}>
+ <div />
+ </InAppAiAgentProvider>,
+ );
+
+ expect(providerMocks.activityUseQuery.mock.calls.at(-1)?.[1]?.enabled).toBe(
+ false,
+ );
+ });
+
it("always allows a persisted background tool for the conversation", async () => {
const approvedSnapshot = {
conversation: {
diff --git a/web/src/features/in-app-agent/components/InAppAgentWindowHost.clienttest.tsx b/web/src/features/in-app-agent/components/InAppAgentWindowHost.clienttest.tsx
index 18055ef292..de813dc264 100644
--- a/web/src/features/in-app-agent/components/InAppAgentWindowHost.clienttest.tsx
+++ b/web/src/features/in-app-agent/components/InAppAgentWindowHost.clienttest.tsx
@@ -8,7 +8,7 @@ const mocks = vi.hoisted(() => ({
}));
vi.mock("@/src/features/in-app-agent/components/InAppAiAgentProvider", () => ({
- useCanUseInAppAgent: () => true,
+ useIsInAppAgentLauncherVisible: () => true,
useInAppAiAgent: () => ({
deleteConversation: vi.fn(),
isExpanded: false,
diff --git a/web/src/features/in-app-agent/components/InAppAgentWindowHost.tsx b/web/src/features/in-app-agent/components/InAppAgentWindowHost.tsx
index f4c626aa8c..0ea0783a02 100644
--- a/web/src/features/in-app-agent/components/InAppAgentWindowHost.tsx
+++ b/web/src/features/in-app-agent/components/InAppAgentWindowHost.tsx
@@ -12,7 +12,7 @@ import {
useInAppAgentWindowShellPanelControl,
} from "@/src/features/in-app-agent/components/InAppAgentWindowShell";
import {
- useCanUseInAppAgent,
+ useIsInAppAgentLauncherVisible,
useInAppAiAgent,
} from "@/src/features/in-app-agent/components/InAppAiAgentProvider";
import { useWatchedPromiseCallback } from "@/src/hooks/useWatchedPromiseCallback";
@@ -64,7 +64,7 @@ function DeleteConversationDialog({
* window unmounts and its geometry resets on every navigation.
*/
export function InAppAgentWindowHost() {
- const canUseAgent = useCanUseInAppAgent();
+ const isInAppAgentLauncherVisible = useIsInAppAgentLauncherVisible();
const { deleteConversation, open, setOpen, isExpanded, setIsExpanded } =
useInAppAiAgent();
const panelRef = useRef<HTMLDivElement>(null);
@@ -121,9 +121,9 @@ export function InAppAgentWindowHost() {
floatingPanelHandle.initializeGeometry();
}, [floatingPanelHandle, isExpanded, open]);
- // Only `canUseAgent` gates the tree: the shell owns the `open` guard so the
- // handheld drawer stays mounted and can animate itself closed.
- if (!canUseAgent) {
+ // Only `isInAppAgentLauncherVisible` gates the tree: the shell owns the `open` guard
+ // so the handheld drawer stays mounted and can animate itself closed.
+ if (!isInAppAgentLauncherVisible) {
return null;
}
diff --git a/web/src/features/in-app-agent/components/InAppAiAgentProvider.tsx b/web/src/features/in-app-agent/components/InAppAiAgentProvider.tsx
index 9b5d96a962..c31ca99378 100644
--- a/web/src/features/in-app-agent/components/InAppAiAgentProvider.tsx
+++ b/web/src/features/in-app-agent/components/InAppAiAgentProvider.tsx
@@ -54,7 +54,6 @@ import {
import type { InAppAgentError } from "@/src/features/in-app-agent/components/utils/utils";
import { useHasEntitlement } from "@/src/features/entitlements/hooks";
import { showErrorToast } from "@/src/features/notifications/showErrorToast";
-import { useLangfuseCloudRegion } from "@/src/features/organizations/hooks";
import { useQueryProjectOrOrganization } from "@/src/features/projects/hooks";
import { api } from "@/src/utils/api";
import {
@@ -293,7 +292,7 @@ function InAppAiAgentProviderInner({
const utils = api.useUtils();
const capture = usePostHogClientCapture();
const session = useSession();
- const isInAppAgentEnabled = useIsInAppAgentEnabled();
+ const canUseAssistant = useCanUseInAppAgent();
const { organization } = useQueryProjectOrOrganization();
const [enableDialogOpen, setEnableDialogOpen] = useState(false);
const [_selectedConversationId, setSelectedConversationId] =
@@ -435,7 +434,7 @@ function InAppAiAgentProviderInner({
userId,
// Polling a project the server will reject turns every page load and
// window focus into a Forbidden toast.
- enabled: isInAppAgentEnabled,
+ enabled: canUseAssistant,
// Only what the user can actually see counts as looked at; a selected
// conversation behind a closed window has not been read.
visibleConversationId: open ? selectedConversationId : null,
@@ -1591,34 +1590,39 @@ export function useInAppAiAgent() {
return ctx;
}
-/** Client mirror of the server's assertInAppAgentAvailable: whether a request
- * for this project would be served, so callers can skip ones it would reject. */
-export function useIsInAppAgentEnabled() {
+/** Whether a run would be admitted for this project. Client mirror of
+ * assertInAppAgentAvailable: instance switch, entitlement, and org AI
+ * Features. Use this for polling and other paths that must not hit the
+ * server when org AI Features is off. More restrictive than
+ * useIsInAppAgentLauncherVisible. */
+export function useCanUseInAppAgent() {
const hasInAppAgentEntitlement = useHasEntitlement("in-app-agent");
- const { isLangfuseCloud } = useLangfuseCloudRegion();
const { organization } = useQueryProjectOrOrganization();
+ const session = useSession();
+ const instanceEnabled = session.data?.environment.inAppAgentEnabled ?? false;
return (
- isLangfuseCloud &&
+ instanceEnabled &&
hasInAppAgentEntitlement &&
Boolean(organization?.aiFeaturesEnabled)
);
}
-/** Whether the current user/context may use the in-app assistant at all. Shared
- * gate for the launcher button and the window host. Deliberately looser than
- * useIsInAppAgentEnabled: with the org AI toggle off the entry points still
- * show, and clicking one opens the dialog that turns it on. */
-export function useCanUseInAppAgent() {
+/** Whether to show Assistant entry points (nav launcher, v4/eval CTAs).
+ * Looser than useCanUseInAppAgent: with org AI Features off they still
+ * show, and clicking opens the dialog that turns it on. Hidden entirely
+ * when the instance-wide switch is off. */
+export function useIsInAppAgentLauncherVisible() {
const { isAvailable } = useInAppAiAgent();
const hasInAppAgentEntitlement = useHasEntitlement("in-app-agent");
- const { isLangfuseCloud } = useLangfuseCloudRegion();
const { organization } = useQueryProjectOrOrganization();
+ const session = useSession();
+ const instanceEnabled = session.data?.environment.inAppAgentEnabled ?? false;
return (
+ instanceEnabled &&
isAvailable &&
hasInAppAgentEntitlement &&
- isLangfuseCloud &&
Boolean(organization)
);
}
diff --git a/web/src/features/in-app-agent/server/availability.ts b/web/src/features/in-app-agent/server/availability.ts
index 9544b74949..0daa31993e 100644
--- a/web/src/features/in-app-agent/server/availability.ts
+++ b/web/src/features/in-app-agent/server/availability.ts
@@ -2,8 +2,11 @@ import type { Session } from "next-auth";
import { BaseError, ForbiddenError } from "@langfuse/shared";
import type { PrismaClient } from "@langfuse/shared/src/db";
+import {
+ getInAppAgentModelConfig,
+ isInAppAgentInstanceEnabled,
+} from "@langfuse/shared/in-app-agent/server/modelProvider";
-import { env } from "@/src/env.mjs";
import { hasEntitlement } from "@/src/features/entitlements/server/hasEntitlement";
export async function assertInAppAgentAvailable({
@@ -15,12 +18,11 @@ export async function assertInAppAgentAvailable({
projectId: string;
user: NonNullable<Session["user"]>;
}) {
- // TODO(LFE-14555): Remove this guard once the OSS release strategy is ready.
- if (!env.NEXT_PUBLIC_LANGFUSE_CLOUD_REGION) {
+ if (!isInAppAgentInstanceEnabled()) {
throw new BaseError(
"PreconditionFailedError",
412,
- "In-app agent is not available in this environment yet.",
+ "In-app agent is not enabled on this instance.",
true,
);
}
@@ -52,8 +54,22 @@ export async function assertInAppAgentAvailable({
});
if (!project?.organization.aiFeaturesEnabled) {
- throw new ForbiddenError("Assistant is not enabled for this organization");
+ throw new ForbiddenError(
+ "In-app agent is not enabled for this organization",
+ );
}
return project.organization;
}
+
+/** Only startRun and approval continuation dispatch to the model. */
+export function assertInAppAgentModelConfigured() {
+ if (!getInAppAgentModelConfig()) {
+ throw new BaseError(
+ "PreconditionFailedError",
+ 412,
+ "In-app agent Bedrock model is not configured. Set LANGFUSE_AWS_BEDROCK_MODEL and LANGFUSE_AWS_BEDROCK_REGION.",
+ true,
+ );
+ }
+}
diff --git a/web/src/features/in-app-agent/server/router.ts b/web/src/features/in-app-agent/server/router.ts
index da0a695968..92a2a40e05 100644
--- a/web/src/features/in-app-agent/server/router.ts
+++ b/web/src/features/in-app-agent/server/router.ts
@@ -18,7 +18,10 @@ import {
getInAppAgentInstrumentationTraceId,
} from "@langfuse/shared/in-app-agent";
import { InAppAgentMessageFeedbackValueSchema } from "../schema";
-import { assertInAppAgentAvailable } from "@/src/features/in-app-agent/server/availability";
+import {
+ assertInAppAgentAvailable,
+ assertInAppAgentModelConfigured,
+} from "@/src/features/in-app-agent/server/availability";
import {
createTRPCRouter,
protectedProjectProcedure,
@@ -209,6 +212,7 @@ export const inAppAgentRouter = createTRPCRouter({
projectId: input.projectId,
user: ctx.session.user,
});
+ assertInAppAgentModelConfigured();
const rateLimitScope = getInAppAgentApiAccessScope(
ctx.session.user,
@@ -268,6 +272,7 @@ export const inAppAgentRouter = createTRPCRouter({
projectId: input.projectId,
user: ctx.session.user,
});
+ assertInAppAgentModelConfigured();
const rateLimitScope = getInAppAgentApiAccessScope(
ctx.session.user,
diff --git a/web/src/features/organizations/components/AIFeatureSwitch.tsx b/web/src/features/organizations/components/AIFeatureSwitch.tsx
index ebd76261cc..72d1efd02d 100644
--- a/web/src/features/organizations/components/AIFeatureSwitch.tsx
+++ b/web/src/features/organizations/components/AIFeatureSwitch.tsx
@@ -114,8 +114,6 @@ export default function AIFeatureSwitch() {
});
}
- if (!isLangfuseCloud) return null;
-
return (
<div>
<Header title="AI Features" />
@@ -125,22 +123,30 @@ export default function AIFeatureSwitch() {
<h4 className="font-bold">
Enable AI powered features for your organization
</h4>
- <p className="text-sm">
- This setting applies to all users and projects. Any data{" "}
- <i>can</i> be sent to AWS Bedrock within the Langfuse data region.
- Traces are sent to Langfuse Cloud in your data region. Your data
- will not be used for training models. Applicable HIPAA, SOC2,
- GDPR, and ISO 27001 compliance remains intact.{" "}
- <a
- href="https://langfuse.com/security/ai-features"
- target="_blank"
- rel="noopener noreferrer"
- className="text-primary inline-flex items-center gap-1 hover:underline"
- >
- More details in the docs here.
- <ExternalLink className="h-3 w-3" />
- </a>
- </p>
+ {isLangfuseCloud ? (
+ <p className="text-sm">
+ This setting applies to all users and projects. Any data{" "}
+ <i>can</i> be sent to AWS Bedrock within the Langfuse data
+ region. Traces are sent to Langfuse Cloud in your data region.
+ Your data will not be used for training models. Applicable
+ HIPAA, SOC2, GDPR, and ISO 27001 compliance remains intact.{" "}
+ <a
+ href="https://langfuse.com/security/ai-features"
+ target="_blank"
+ rel="noopener noreferrer"
+ className="text-primary inline-flex items-center gap-1 hover:underline"
+ >
+ More details in the docs here.
+ <ExternalLink className="h-3 w-3" />
+ </a>
+ </p>
+ ) : (
+ <p className="text-sm">
+ This setting applies to all users and projects. When enabled,
+ the assistant can send relevant project data to the model
+ provider configured by your instance administrator.
+ </p>
+ )}
</div>
<div className="relative">
<Switch
@@ -155,7 +161,7 @@ export default function AIFeatureSwitch() {
)}
</div>
</div>
- {isAIFeatureSwitchEnabled && (
+ {isLangfuseCloud && isAIFeatureSwitchEnabled && (
<div className="mt-4 flex flex-row items-center justify-between border-t pt-4">
<div className="flex flex-col gap-1">
<h4 className="font-bold">
@@ -200,20 +206,27 @@ export default function AIFeatureSwitch() {
<strong>
{isAIFeatureSwitchEnabled ? "enable " : "disable"}
</strong>{" "}
- AI features for your organization. When enabled, any data{" "}
- <i>can</i> be sent to AWS Bedrock in your data region for
- processing.
- <br />
- <br />{" "}
- <a
- href="https://langfuse.com/security/ai-features"
- target="_blank"
- rel="noopener noreferrer"
- className="text-primary inline-flex items-center gap-1 hover:underline"
- >
- Learn more in the docs.
- <ExternalLink className="h-3 w-3" />
- </a>
+ AI features for your organization. When enabled, relevant data can
+ be sent{" "}
+ {isLangfuseCloud
+ ? "to AWS Bedrock in your data region"
+ : "to the model provider configured by your instance administrator"}{" "}
+ for processing.
+ {isLangfuseCloud && (
+ <>
+ <br />
+ <br />{" "}
+ <a
+ href="https://langfuse.com/security/ai-features"
+ target="_blank"
+ rel="noopener noreferrer"
+ className="text-primary inline-flex items-center gap-1 hover:underline"
+ >
+ Learn more in the docs.
+ <ExternalLink className="h-3 w-3" />
+ </a>
+ </>
+ )}
</span>
<p className="text-muted-foreground mt-3 text-sm">
Are you sure you want to proceed?
diff --git a/web/src/features/organizations/server/organizationRouter.ts b/web/src/features/organizations/server/organizationRouter.ts
index fc77a204e6..9c457f296c 100644
--- a/web/src/features/organizations/server/organizationRouter.ts
+++ b/web/src/features/organizations/server/organizationRouter.ts
@@ -194,13 +194,13 @@ export const organizationsRouter = createTRPCRouter({
});
if (
- (input.aiFeaturesEnabled !== undefined ||
- input.aiTelemetryEnabled !== undefined) &&
+ input.aiTelemetryEnabled !== undefined &&
!env.NEXT_PUBLIC_LANGFUSE_CLOUD_REGION
) {
throw new TRPCError({
code: "PRECONDITION_FAILED",
- message: "AI features are not available in self-hosted deployments.",
+ message:
+ "AI telemetry controls are only available on Langfuse Cloud.",
});
}
diff --git a/web/src/features/v4-migration/EvaluatorMigrationDialog.tsx b/web/src/features/v4-migration/EvaluatorMigrationDialog.tsx
index c26d682266..511e5279b0 100644
--- a/web/src/features/v4-migration/EvaluatorMigrationDialog.tsx
+++ b/web/src/features/v4-migration/EvaluatorMigrationDialog.tsx
@@ -11,7 +11,7 @@ import {
import { Button } from "@/src/components/ui/button";
import { CodeBlock } from "@/src/components/design-system/Codeblock/Codeblock";
import {
- useCanUseInAppAgent,
+ useIsInAppAgentLauncherVisible,
useInAppAiAgent,
} from "@/src/features/in-app-agent/components/InAppAiAgentProvider";
import { usePostHogClientCapture } from "@/src/features/posthog-analytics/usePostHogClientCapture";
@@ -47,7 +47,7 @@ export function EvaluatorMigrationDialog({
onAssistantStarted,
initialAction,
}: EvaluatorMigrationDialogProps) {
- const canUseAssistant = useCanUseInAppAgent();
+ const isInAppAgentLauncherVisible = useIsInAppAgentLauncherVisible();
const { organization } = useQueryProjectOrOrganization();
const { openAssistant, submit } = useInAppAiAgent();
const canUpdateOrgSettings = useHasOrganizationAccess({
@@ -60,7 +60,7 @@ export function EvaluatorMigrationDialog({
const aiFeaturesEnabled = Boolean(organization?.aiFeaturesEnabled);
const isSingleEvaluator = scope.type === "single";
- const showAssistantOption = canUseAssistant;
+ const showAssistantOption = isInAppAgentLauncherVisible;
// With AI features disabled, the choice screen's assistant option owns the
// enable-AI and admin-handoff side effects, so the preselect only applies
// once the assistant can actually start.
diff --git a/web/src/features/v4-migration/V4MigrationContent.clienttest.tsx b/web/src/features/v4-migration/V4MigrationContent.clienttest.tsx
index a1ed32e399..0e95dddaea 100644
--- a/web/src/features/v4-migration/V4MigrationContent.clienttest.tsx
+++ b/web/src/features/v4-migration/V4MigrationContent.clienttest.tsx
@@ -226,7 +226,7 @@ vi.mock("@/src/features/v4-migration/useV4UpgradeAssistantSupport", () => ({
}));
vi.mock("@/src/features/in-app-agent/components/InAppAiAgentProvider", () => ({
- useCanUseInAppAgent: () => true,
+ useIsInAppAgentLauncherVisible: () => true,
useInAppAiAgent: () => ({
openAssistant: mocks.openAssistant,
submit: mocks.submitAgentMessage,
diff --git a/web/src/features/v4-migration/V4MigrationContent.tsx b/web/src/features/v4-migration/V4MigrationContent.tsx
index c9381e62f5..572967d785 100644
--- a/web/src/features/v4-migration/V4MigrationContent.tsx
+++ b/web/src/features/v4-migration/V4MigrationContent.tsx
@@ -10,7 +10,7 @@ import {
Info,
} from "lucide-react";
import { env } from "@/src/env.mjs";
-import { useCanUseInAppAgent } from "@/src/features/in-app-agent/components/InAppAiAgentProvider";
+import { useIsInAppAgentLauncherVisible } from "@/src/features/in-app-agent/components/InAppAiAgentProvider";
import { useSupportDrawer } from "@/src/features/support-chat/SupportDrawerProvider";
import { Button } from "@/src/components/ui/button";
import {
@@ -1401,7 +1401,7 @@ export function V4MigrationDetailsContent({
onNavigate?.();
openSupportDrawerWithMode("form", { topic: "V4 Migration" });
};
- const canUseAssistant = useCanUseInAppAgent();
+ const isInAppAgentLauncherVisible = useIsInAppAgentLauncherVisible();
// Same org source as EvaluatorMigrationDialog's gating, so the button
// branding and the dialog's preselect can't disagree.
const { organization: routeOrganization } = useQueryProjectOrOrganization();
@@ -1461,7 +1461,7 @@ export function V4MigrationDetailsContent({
<V4MigrationEvalsSection
state={migrationData.evals}
assistant={
- canUseAssistant
+ isInAppAgentLauncherVisible
? {
onMigrate: handleMigrateEvalsWithAgent,
aiFeaturesEnabled,
diff --git a/web/src/features/v4-migration/V4MigrationDelayBadge.clienttest.tsx b/web/src/features/v4-migration/V4MigrationDelayBadge.clienttest.tsx
index a1b42a26fc..06ae234eea 100644
--- a/web/src/features/v4-migration/V4MigrationDelayBadge.clienttest.tsx
+++ b/web/src/features/v4-migration/V4MigrationDelayBadge.clienttest.tsx
@@ -161,7 +161,7 @@ vi.mock("@/src/features/v4-migration/hooks/useV4MigrationData", () => ({
}));
vi.mock("@/src/features/in-app-agent/components/InAppAiAgentProvider", () => ({
- useCanUseInAppAgent: () => false,
+ useIsInAppAgentLauncherVisible: () => false,
useInAppAiAgent: () => ({ setOpen: vi.fn(), submit: vi.fn() }),
}));
diff --git a/web/src/features/v4-migration/useV4UpgradeAssistantSupport.clienttest.ts b/web/src/features/v4-migration/useV4UpgradeAssistantSupport.clienttest.ts
index 6436f6efa1..cbd794e037 100644
--- a/web/src/features/v4-migration/useV4UpgradeAssistantSupport.clienttest.ts
+++ b/web/src/features/v4-migration/useV4UpgradeAssistantSupport.clienttest.ts
@@ -5,7 +5,7 @@ import { describe, expect, it, vi } from "vitest";
import { V4_CODING_AGENT_PROMPT } from "./useV4UpgradeAssistantSupport";
vi.mock("@/src/features/in-app-agent/components/InAppAiAgentProvider", () => ({
- useCanUseInAppAgent: () => true,
+ useIsInAppAgentLauncherVisible: () => true,
}));
describe("V4_CODING_AGENT_PROMPT", () => {
diff --git a/web/src/features/v4-migration/useV4UpgradeAssistantSupport.ts b/web/src/features/v4-migration/useV4UpgradeAssistantSupport.ts
index 79afb8ef60..4bcb953318 100644
--- a/web/src/features/v4-migration/useV4UpgradeAssistantSupport.ts
+++ b/web/src/features/v4-migration/useV4UpgradeAssistantSupport.ts
@@ -1,4 +1,4 @@
-import { useCanUseInAppAgent } from "@/src/features/in-app-agent/components/InAppAiAgentProvider";
+import { useIsInAppAgentLauncherVisible } from "@/src/features/in-app-agent/components/InAppAiAgentProvider";
import { useProjectV4SdkData } from "@/src/features/v4-migration/hooks/useV4MigrationData";
import { api } from "@/src/utils/api";
@@ -57,7 +57,7 @@ export function useEvalUpgradeAssistantPlan(params: {
projectId: string | undefined;
enabled: boolean;
}) {
- const canUseAssistant = useCanUseInAppAgent();
+ const isInAppAgentLauncherVisible = useIsInAppAgentLauncherVisible();
const sdk = useProjectV4SdkData(params);
const evalQuery = api.v4Transition.traceLevelEvalSummary.useQuery(
{ projectId: params.projectId ?? "" },
@@ -77,10 +77,10 @@ export function useEvalUpgradeAssistantPlan(params: {
: "outside";
return {
- canUseAssistant,
+ canUseAssistant: isInAppAgentLauncherVisible,
mode,
/** Whether the migration panel may show the Assistant CTA. */
- showAssistantButton: canUseAssistant && mode !== "outside",
+ showAssistantButton: isInAppAgentLauncherVisible && mode !== "outside",
assistantPrompt:
mode === "evals-ready"
? EVAL_UPGRADE_ASSISTANT_PROMPT
diff --git a/web/src/server/auth.ts b/web/src/server/auth.ts
index 281633bd5e..0188cff64c 100644
--- a/web/src/server/auth.ts
+++ b/web/src/server/auth.ts
@@ -6,6 +6,7 @@ import {
} from "next-auth";
import { PrismaAdapter } from "@next-auth/prisma-adapter";
import { prisma } from "@langfuse/shared/src/db";
+import { isInAppAgentInstanceEnabled } from "@langfuse/shared/in-app-agent/server/modelProvider";
import {
hashPassword,
verifyPassword,
@@ -879,6 +880,7 @@ export async function getAuthOptions(signupAttribution?: {
environment: {
enableExperimentalFeatures:
env.LANGFUSE_ENABLE_EXPERIMENTAL_FEATURES === "true",
+ inAppAgentEnabled: isInAppAgentInstanceEnabled(),
// Enables features that are only available under an enterprise license when self-hosting Langfuse
// If you edit this line, you risk executing code that is not MIT licensed (self-contained in /ee folders otherwise)
selfHostedInstancePlan: getSelfHostedInstancePlanServerSide(),
diff --git a/web/types/next-auth.d.ts b/web/types/next-auth.d.ts
index 24bb2b71ea..5b3ccd2f57 100644
--- a/web/types/next-auth.d.ts
+++ b/web/types/next-auth.d.ts
@@ -57,6 +57,9 @@ declare module "next-auth" {
environment: {
// Run-time environment variables that need to be available client-side
enableExperimentalFeatures: boolean;
+ // Instance-wide in-app agent switch. Populated by the session callback.
+ // Optional so existing session mocks need not set it.
+ inAppAgentEnabled?: boolean;
// Enables features that are only available under an enterprise/commercial license when self-hosting Langfuse
selfHostedInstancePlan: Plan | null;
// V4 migration write mode. Mirrors LANGFUSE_MIGRATION_V4_WRITE_MODE so the
diff --git a/worker/package.json b/worker/package.json
index b13fa6b79f..87456b4ca2 100644
--- a/worker/package.json
+++ b/worker/package.json
@@ -37,7 +37,6 @@
"@appsignal/opentelemetry-instrumentation-bullmq": "^0.8.0",
"@langfuse/shared": "workspace:*",
"@aws-sdk/client-lambda-microvms": "catalog:",
- "@aws-sdk/credential-providers": "catalog:",
"@mastra/core": "1.46.0",
"@mastra/mcp": "1.8.1",
"@opentelemetry/api": "^1.9.1",
diff --git a/worker/src/features/in-app-agent/executeInAppAgentRun.test.ts b/worker/src/features/in-app-agent/executeInAppAgentRun.test.ts
index 8dd39e241f..2de17464b3 100644
--- a/worker/src/features/in-app-agent/executeInAppAgentRun.test.ts
+++ b/worker/src/features/in-app-agent/executeInAppAgentRun.test.ts
@@ -35,9 +35,13 @@ type AgentScenario = (ctx: {
};
signal: AbortSignal;
options: {
- awsBedrock: {
- profile?: string;
+ model: {
+ provider: "bedrock";
+ modelId: string;
};
+ awsProfile?: string;
+ langfuseClient?: unknown;
+ useLocalPrompt: boolean;
langfuseMcp: {
toolPolicy: {
available: ReadonlySet<string>;
@@ -59,6 +63,7 @@ const scenarioRef = vi.hoisted(() => ({
failApiKeyDelete: false,
apiKeyDeleteCalls: 0,
titleInferenceCalls: 0,
+ instanceEnabled: true,
}));
vi.mock("./runtime/agent", async (importOriginal) => {
@@ -103,6 +108,21 @@ vi.mock(
}),
);
+vi.mock(
+ "@langfuse/shared/in-app-agent/server/modelProvider",
+ async (importOriginal) => {
+ const actual =
+ await importOriginal<
+ typeof import("@langfuse/shared/in-app-agent/server/modelProvider")
+ >();
+
+ return {
+ ...actual,
+ isInAppAgentInstanceEnabled: () => scenarioRef.instanceEnabled,
+ };
+ },
+);
+
vi.mock("@langfuse/shared/src/server/auth/apiKeys", async (importOriginal) => {
const actual =
await importOriginal<
@@ -271,6 +291,7 @@ async function seedApprovedContinuation(opts?: {
describe("executeInAppAgentRun", () => {
beforeEach(() => {
scenarioRef.titleInferenceCalls = 0;
+ scenarioRef.instanceEnabled = true;
});
it("does not regenerate the conversation title after executing a user-message run", async () => {
@@ -296,7 +317,7 @@ describe("executeInAppAgentRun", () => {
const { projectId, run } = await seedBackgroundRun();
scenarioRef.current = async ({ options }) => {
- expect(options.awsBedrock.profile).toBe("developer-profile");
+ expect(options.awsProfile).toBe("developer-profile");
await options.onComplete();
await options.onFinish();
};
@@ -309,6 +330,25 @@ describe("executeInAppAgentRun", () => {
}
});
+ it("uses the bundled prompt in self-hosted production", async () => {
+ const originalCloudRegion = env.NEXT_PUBLIC_LANGFUSE_CLOUD_REGION;
+ env.NEXT_PUBLIC_LANGFUSE_CLOUD_REGION = undefined;
+
+ const { projectId, run } = await seedBackgroundRun();
+ scenarioRef.current = async ({ options }) => {
+ expect(options.useLocalPrompt).toBe(true);
+ expect(options.langfuseClient).toBeUndefined();
+ await options.onComplete();
+ await options.onFinish();
+ };
+
+ try {
+ await executeInAppAgentRun({ projectId, runId: run.id });
+ } finally {
+ env.NEXT_PUBLIC_LANGFUSE_CLOUD_REGION = originalCloudRegion;
+ }
+ });
+
it("passes persisted continuation context to the agent input", async () => {
const rootRunId = "root-run-1";
const traceStartedAt = "2026-08-14T10:00:00.000Z";
@@ -714,6 +754,21 @@ describe("executeInAppAgentRun", () => {
expect(await getInAppAgentApiKeys(projectId)).toHaveLength(0);
});
+ it("fails revalidation at claim as FAILED (init_failed) when in-app agent is instance-disabled", async () => {
+ scenarioRef.instanceEnabled = false;
+ const { projectId, run } = await seedBackgroundRun();
+ scenarioRef.current = async () => {
+ throw new Error("agent loop must not start when the instance is off");
+ };
+
+ await executeInAppAgentRun({ projectId, runId: run.id });
+
+ const failed = await getRun(projectId, run.id);
+ expect(failed.status).toBe("FAILED");
+ expect(failed.errorCode).toBe("init_failed");
+ expect(await getInAppAgentApiKeys(projectId)).toHaveLength(0);
+ });
+
it("keeps the MCP-key pointer when the delete fails, for reconciliation to retry", async () => {
const { projectId, run } = await seedBackgroundRun();
diff --git a/worker/src/features/in-app-agent/executeInAppAgentRun.ts b/worker/src/features/in-app-agent/executeInAppAgentRun.ts
index a712285b11..05dd1d3349 100644
--- a/worker/src/features/in-app-agent/executeInAppAgentRun.ts
+++ b/worker/src/features/in-app-agent/executeInAppAgentRun.ts
@@ -1,6 +1,5 @@
import { Role } from "@langfuse/shared";
import { prisma } from "@langfuse/shared/src/db";
-import { env as sharedEnv } from "@langfuse/shared/src/env";
import {
getLangfuseAITraceSinkParams,
logger,
@@ -29,6 +28,10 @@ import {
toPersistableAgentEvent,
type PersistedConversationEvent,
} from "@langfuse/shared/in-app-agent/server/persistence";
+import {
+ getInAppAgentModelConfig,
+ isInAppAgentInstanceEnabled,
+} from "@langfuse/shared/in-app-agent/server/modelProvider";
import {
claimQueuedRun,
clearRunMcpApiKeyPointer,
@@ -153,20 +156,23 @@ export async function executeInAppAgentRun(params: {
try {
// ---- Revalidate at claim; nothing from enqueue time is trusted. ----
- if (!env.NEXT_PUBLIC_LANGFUSE_CLOUD_REGION) {
+ if (!isInAppAgentInstanceEnabled()) {
throw new InAppAgentRunInitError(
- "In-app agent is only available on Langfuse Cloud",
+ "In-app agent is not enabled on this instance",
);
}
- const bedrockModelId = run.model ?? sharedEnv.LANGFUSE_AWS_BEDROCK_MODEL;
+ const modelConfig = getInAppAgentModelConfig({ modelId: run.model });
- if (!bedrockModelId || !sharedEnv.LANGFUSE_AWS_BEDROCK_REGION) {
+ if (!modelConfig) {
throw new InAppAgentRunInitError(
- "Assistant Bedrock model is not configured",
+ "In-app agent Bedrock model is not configured",
);
}
+ const useBundledPrompt =
+ env.NODE_ENV === "development" || !env.NEXT_PUBLIC_LANGFUSE_CLOUD_REGION;
+
const conversation = await prisma.inAppAgentConversation.findFirst({
where: { id: run.conversationId, projectId, deletedAt: null },
});
@@ -500,11 +506,8 @@ export async function executeInAppAgentRun(params: {
await cleanupMcpApiKeyLogged();
await sandboxState?.onTurnEnded();
},
- awsBedrock: {
- region: sharedEnv.LANGFUSE_AWS_BEDROCK_REGION,
- modelId: bedrockModelId,
- ...(awsProfile ? { profile: awsProfile } : {}),
- },
+ model: modelConfig,
+ ...(awsProfile ? { awsProfile } : {}),
langfuseMcp: {
url: getLangfuseMcpUrl(),
publicKey: mcpApiKey.publicKey,
@@ -516,8 +519,10 @@ export async function executeInAppAgentRun(params: {
projectId,
isV4Enabled: access.v4BetaEnabled,
},
- langfuseClient: getInAppAgentPromptClient(),
- useLocalPrompt: env.NODE_ENV === "development",
+ useLocalPrompt: useBundledPrompt,
+ ...(useBundledPrompt
+ ? {}
+ : { langfuseClient: getInAppAgentPromptClient() }),
langfuseTracing: buildTracingConfig({
aiTelemetryEnabled: project.organization.aiTelemetryEnabled,
projectId,
diff --git a/worker/src/features/in-app-agent/runtime/agent.test.ts b/worker/src/features/in-app-agent/runtime/agent.test.ts
index 3207df5ee8..9328390ab2 100644
--- a/worker/src/features/in-app-agent/runtime/agent.test.ts
+++ b/worker/src/features/in-app-agent/runtime/agent.test.ts
@@ -25,6 +25,13 @@ import type { InAppAgentTracingConfig } from "./instrumentation";
const EXPECTED_MCP_USER_AGENT = "langfuse-in-app-agent";
+const testBedrockModel = (modelId: string) => ({
+ provider: "bedrock" as const,
+ modelId,
+ titleModelId: modelId,
+ region: "eu-central-1",
+});
+
// Shape of the tool entries the mocked MCP client feeds into the Agent
// constructor. `Agent`'s own `tools` type is a `DynamicArgument` union that
// does not allow property access, so tests read it through this view.
@@ -472,7 +479,10 @@ describe("createAgUiStream", () => {
});
});
- const initializeBasicTracedAgent = async (runId: string) => {
+ const initializeBasicTracedAgent = async (
+ runId: string,
+ model = testBedrockModel("test-model"),
+ ) => {
const { createAgUiStream } = await import("./agent");
const input = {
threadId: "conversation-1",
@@ -506,7 +516,7 @@ describe("createAgUiStream", () => {
input,
signal: new AbortController().signal,
options: {
- awsBedrock: { modelId: "test-model" },
+ model,
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -524,6 +534,18 @@ describe("createAgUiStream", () => {
await readStream(stream);
};
+ it("uses the shared Bedrock default-credential auth", async () => {
+ const { createAmazonBedrock } = await import("ai-sdk-amazon-bedrock-v4");
+
+ await initializeBasicTracedAgent("run-default-bedrock-auth");
+
+ expect(createAmazonBedrock).toHaveBeenCalledWith({
+ region: "eu-central-1",
+ apiKey: "",
+ credentialProvider: expect.any(Function),
+ });
+ });
+
it("forwards model stream parts when finish tracing throws", async () => {
instrumentationMocks.instrumentation.recordModelStreamPart
.mockImplementationOnce(() => undefined)
@@ -685,9 +707,7 @@ describe("createAgUiStream", () => {
eventOrder.push(`persist:${event.type}`);
await Promise.resolve();
},
- awsBedrock: {
- modelId: "eu.anthropic.claude-opus-4-8",
- },
+ model: testBedrockModel("eu.anthropic.claude-opus-4-8"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -973,9 +993,7 @@ describe("createAgUiStream", () => {
input,
signal: new AbortController().signal,
options: {
- awsBedrock: {
- modelId: "meta.llama3-70b-instruct-v1:0",
- },
+ model: testBedrockModel("meta.llama3-70b-instruct-v1:0"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -1046,7 +1064,7 @@ describe("createAgUiStream", () => {
input,
signal: new AbortController().signal,
options: {
- awsBedrock: { modelId: "eu.anthropic.claude-opus-4-8" },
+ model: testBedrockModel("eu.anthropic.claude-opus-4-8"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -1143,7 +1161,7 @@ describe("createAgUiStream", () => {
onEvent: (event) => {
persistedEvents.push(event);
},
- awsBedrock: { modelId: "test-model" },
+ model: testBedrockModel("test-model"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -1215,7 +1233,7 @@ describe("createAgUiStream", () => {
input,
signal: new AbortController().signal,
options: {
- awsBedrock: { modelId: "eu.anthropic.claude-opus-4-8" },
+ model: testBedrockModel("eu.anthropic.claude-opus-4-8"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -1263,7 +1281,7 @@ describe("createAgUiStream", () => {
onEvent: (event) => {
persistedEvents.push(event);
},
- awsBedrock: { modelId: "test-model" },
+ model: testBedrockModel("test-model"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -1550,7 +1568,7 @@ describe("createAgUiStream", () => {
onEvent: (event) => {
persistedEvents.push(event);
},
- awsBedrock: { modelId: "test-model" },
+ model: testBedrockModel("test-model"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -1627,7 +1645,7 @@ describe("createAgUiStream", () => {
persistedEvents.push(event);
},
onError,
- awsBedrock: { modelId: "test-model" },
+ model: testBedrockModel("test-model"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -1791,7 +1809,7 @@ describe("createAgUiStream", () => {
input,
signal: new AbortController().signal,
options: {
- awsBedrock: { modelId: "test-model" },
+ model: testBedrockModel("test-model"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -1872,7 +1890,7 @@ describe("createAgUiStream", () => {
persistedEvents.push(event);
},
onComplete,
- awsBedrock: { modelId: "test-model" },
+ model: testBedrockModel("test-model"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -2021,7 +2039,7 @@ describe("createAgUiStream", () => {
input,
signal: new AbortController().signal,
options: {
- awsBedrock: { modelId: "test-model" },
+ model: testBedrockModel("test-model"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
@@ -2109,7 +2127,7 @@ describe("createAgUiStream", () => {
input,
signal: new AbortController().signal,
options: {
- awsBedrock: { modelId: "test-model" },
+ model: testBedrockModel("test-model"),
langfuseMcp: {
url: "https://example.com/api/public/mcp",
publicKey: "pk",
diff --git a/worker/src/features/in-app-agent/runtime/agent.ts b/worker/src/features/in-app-agent/runtime/agent.ts
index deb8cd5c80..2291dd1527 100644
--- a/worker/src/features/in-app-agent/runtime/agent.ts
+++ b/worker/src/features/in-app-agent/runtime/agent.ts
@@ -2,7 +2,6 @@ import { EventType } from "@ag-ui/core";
import { MastraAgent } from "@ag-ui/mastra";
import { IN_APP_AGENT_SYSTEM_PROMPT_TEMPLATE } from "@langfuse/shared/in-app-agent/server/systemPrompt";
import { createAmazonBedrock } from "ai-sdk-amazon-bedrock-v4";
-import { fromNodeProviderChain } from "@aws-sdk/credential-providers";
import { Agent } from "@mastra/core/agent";
import { MCPClient } from "@mastra/mcp";
import type { Langfuse } from "langfuse";
@@ -45,11 +44,15 @@ import {
import { LANGFUSE_IN_APP_AGENT_SKILLS } from "./skills";
import type { InAppAgentSandbox } from "./sandbox";
import { DEFAULT_SIDEBAR_HIDDEN_ENVIRONMENTS } from "@langfuse/shared";
-import { logger } from "@langfuse/shared/src/server";
+import {
+ createDefaultBedrockProviderAuth,
+ logger,
+} from "@langfuse/shared/src/server";
import {
IN_APP_AGENT_MCP_TOOL_OVERRIDE_HEADER,
IN_APP_AGENT_REDIRECT_TOOL_NAME,
} from "@langfuse/shared/in-app-agent";
+import type { InAppAgentModelConfig } from "@langfuse/shared/in-app-agent/server/modelProvider";
const ASSISTANT_TITLE = "Langfuse Assistant";
const IN_APP_AGENT_SYSTEM_PROMPT_NAME = "in-app-agent-system-prompt";
@@ -182,11 +185,8 @@ type CreateAgUiStreamOptions = {
onAbort?: () => void | Promise<void>;
onError?: (error: unknown) => void | Promise<void>;
onFinish?: () => void | Promise<void>;
- awsBedrock: {
- region?: string;
- profile?: string;
- modelId: string;
- };
+ model: InAppAgentModelConfig;
+ awsProfile?: string;
langfuseMcp: {
url: string;
publicKey: string;
@@ -198,7 +198,7 @@ type CreateAgUiStreamOptions = {
projectId: string;
isV4Enabled: boolean;
};
- langfuseClient: Langfuse;
+ langfuseClient?: Langfuse;
useLocalPrompt: boolean;
langfuseTracing?: InAppAgentTracingConfig;
sandbox?: InAppAgentSandbox;
@@ -212,7 +212,7 @@ export async function createAgUiStream(params: {
options: CreateAgUiStreamOptions;
}) {
const encoder = new TextEncoder();
- const awsProfile = params.options.awsBedrock.profile;
+ const awsProfile = params.options.awsProfile;
const langfuseMcpAuthHeader = `Basic ${Buffer.from(
`${params.options.langfuseMcp.publicKey}:${params.options.langfuseMcp.secretKey}`,
@@ -236,7 +236,7 @@ export async function createAgUiStream(params: {
tracing: params.options.langfuseTracing
? { ...params.options.langfuseTracing, prompt }
: undefined,
- model: params.options.awsBedrock.modelId,
+ model: params.options.model.modelId,
});
const recordInstrumentation = (
operation: string,
@@ -846,15 +846,17 @@ type ExecutableInAppAgentTool = {
toModelOutput?: (output: unknown) => unknown | PromiseLike<unknown>;
};
-type BedrockLanguageModel = ReturnType<ReturnType<typeof createAmazonBedrock>>;
+type InAppAgentLanguageModel = ReturnType<
+ ReturnType<typeof createAmazonBedrock>
+>;
function withModelTracing(
- model: BedrockLanguageModel,
+ model: InAppAgentLanguageModel,
callbacks: {
onStart?: (options: unknown) => void;
onStreamPart?: (part: unknown) => void;
},
-): BedrockLanguageModel {
+): InAppAgentLanguageModel {
if (!callbacks.onStart && !callbacks.onStreamPart) {
return model;
}
@@ -941,11 +943,9 @@ async function createMastraAdapter(params: {
onToolExecutionEnd?: (toolCallId: string) => void;
}) {
const bedrock = createAmazonBedrock({
- ...(params.options.awsBedrock.region
- ? { region: params.options.awsBedrock.region }
- : {}),
- credentialProvider: fromNodeProviderChain(
- params.awsProfile ? { profile: params.awsProfile } : {},
+ region: params.options.model.region,
+ ...createDefaultBedrockProviderAuth(
+ params.awsProfile ? { profile: params.awsProfile } : undefined,
),
});
@@ -1061,7 +1061,7 @@ async function createMastraAdapter(params: {
params.onToolsAvailable?.(tools);
const reasoningProviderOptions = getBedrockReasoningProviderOptions(
- params.options.awsBedrock.modelId,
+ params.options.model.modelId,
);
// @ag-ui/mastra currently forwards only assistant, user, and tool
@@ -1071,9 +1071,7 @@ async function createMastraAdapter(params: {
// guidance on resumed runs.
let developerGuidance: string | undefined;
const model = withModelTracing(
- bedrock(
- params.options.awsBedrock.modelId as Parameters<typeof bedrock>[0],
- ),
+ bedrock(params.options.model.modelId as Parameters<typeof bedrock>[0]),
{
onStart: params.onModelCallStart,
onStreamPart: params.onModelStreamPart,
@@ -1373,7 +1371,7 @@ function getToolErrorMessage(chunk: MastraApprovalStreamChunk): string {
}
async function getSystemPromptInstructions(params: {
- langfuseClient: Langfuse;
+ langfuseClient?: Langfuse;
useLocalPrompt: boolean;
variables: {
currentDate: string;
@@ -1397,6 +1395,10 @@ async function getSystemPromptInstructions(params: {
};
}
+ if (!params.langfuseClient) {
+ throw new Error("Managed in-app agent prompt client is not configured");
+ }
+
const prompt = await params.langfuseClient.getPrompt(
IN_APP_AGENT_SYSTEM_PROMPT_NAME,
undefined,
@@ -1458,7 +1460,7 @@ function createRunErrorEvent(
error: unknown,
): AgUiEvent {
const message =
- error instanceof Error ? error.message : "Unknown assistant error";
+ error instanceof Error ? error.message : "Unknown in-app agent error";
return {
type: EventType.RUN_ERROR,
@@ -1471,5 +1473,5 @@ function createRunErrorEvent(
function getRunErrorMessage(event: AgUiEvent) {
return typeof event.message === "string" && event.message.trim()
? event.message
- : "Unknown assistant error";
+ : "Unknown in-app agent error";
}
diff --git a/worker/src/features/in-app-agent/runtime/instrumentation.ts b/worker/src/features/in-app-agent/runtime/instrumentation.ts
index 114373d8e0..0cfecf2e53 100644
--- a/worker/src/features/in-app-agent/runtime/instrumentation.ts
+++ b/worker/src/features/in-app-agent/runtime/instrumentation.ts
@@ -536,7 +536,7 @@ export class InAppAgentInstrumentation {
this.endWithError(
typeof event.message === "string"
? event.message
- : "Unknown assistant error",
+ : "Unknown in-app agent error",
);
return;
}