Skip to main content
Belvedir auto-instruments these libraries once initialize() has run. The lists differ by language, so check the one for your SDK: Node (belvedir on npm)
  • OpenAI (Azure OpenAI works through the same openai client)
  • Anthropic
  • Cohere
  • Amazon Bedrock
  • Google Vertex AI and Google Generative AI (Gemini)
  • Together
  • Frameworks: LangChain, LlamaIndex, MCP
  • Vector stores: Chroma, Pinecone, Qdrant
Replicate and HuggingFace are not instrumented on Node; call them through an OpenAI-compatible endpoint (below) or emit spans yourself. Python (belvedir on PyPI)
  • OpenAI (Azure OpenAI works through the same openai client)
  • Anthropic
  • Cohere
  • Amazon Bedrock and SageMaker
  • Google Vertex AI and Google Generative AI (Gemini)
  • Mistral, Groq, Together, Replicate, Ollama, IBM watsonx, Aleph Alpha, Writer, Voyage AI
  • HuggingFace Transformers
  • Frameworks: LangChain, LlamaIndex, Haystack, CrewAI, LiteLLM, OpenAI Agents, Agno, MCP
  • Vector stores: Chroma, Pinecone, Qdrant, Weaviate, Milvus, LanceDB, Marqo

​
OpenAI-compatible endpoints

Local models and hosted OpenAI-compatible endpoints (Together, Fireworks, Groq, OpenRouter, self-hosted vLLM) work through the OpenAI SDK, and raw HTTP calls to any chat/completions endpoint are captured too (see What is captured). Setup is on Local and hosted models; Belvedir’s own hosted models are called the same way. Apps built on the AI SDK (the ai package) are captured only when calls go through @ai-sdk/openai-compatible; see the AI SDK guide. For agents you didn’t write (OpenClaw, Hermes, or anything that already speaks OpenTelemetry), see Agents without the SDK.

​
What is captured

* Requires the current belvedir release (npm install belvedir@latest / pip install -U belvedir).