Troubleshoot the self-hosted Langfuse Assistant
Use these checks when the self-hosted Langfuse Assistant does not start or work as expected. For the complete configuration flow, see Set up the self-hosted Langfuse Assistant.
The Assistant is not visible
Confirm LANGFUSE_IN_APP_AGENT_ENABLED=true on both web and worker, then recreate both services. Enabling organization AI Features does not show the Assistant on its own.
Runs stay queued or time out
Confirm at least one worker is running with LANGFUSE_IN_APP_AGENT_ENABLED=true and shares Redis with web. A worker without it does not consume the run queue.
Runs fail during agent initialization
The worker could not connect to the MCP server. Confirm it can reach /api/public/mcp at LANGFUSE_MCP_BASE_URL when set, otherwise at NEXTAUTH_URL. Add a different internal hostname to LANGFUSE_MCP_ALLOWED_HOSTS on web. A 403 with Invalid Host header means the hostname is not allowed.
The UI reports that the model is not configured
Set LANGFUSE_AI_PROVIDER and LANGFUSE_AI_MODEL on both services. Also set LANGFUSE_AI_API_KEY for anthropic or openai, then recreate both services.
Bedrock returns access denied or model not found
Check the Region, model access, and invoke permissions in Configure the model. An AWS Marketplace subscription error means an administrator must activate the model for the account first.
An OpenAI-compatible gateway returns 404
Include /v1 in LANGFUSE_AI_BASE_URL. The endpoint must serve POST {base}/chat/completions, or POST {base}/responses when LANGFUSE_AI_USE_RESPONSES_API=true.
Sandbox tools are missing
This is expected when LANGFUSE_IN_APP_AGENT_SANDBOX_PROVIDER is unset. Configure lambda-microvm and all required sandbox variables to add file and code-execution tools.
File or code-execution tools fail
Confirm lambda:PassNetworkConnector covers HTTP_INGRESS and the egress connector. Also confirm that the MicroVM image matches the running Langfuse version.
The image build fails
Confirm Lambda MicroVMs are available in the Region, the AWS CLI supports lambda-microvms, the build role can read the artifact, and the caller has iam:PassRole on the build role.
Last updated on
