Supported Libraries
Maev does two related but different things:
- Captures telemetry for your agent runs
- Actively intercepts LLM calls for retries, circuit breaking, and Gateway routing on supported agent and provider layers
Those support levels are not identical, so this page separates them clearly.
Direct prompt injection support
These libraries and frameworks support both telemetry capture and active SDK interception with no extra configuration:
| Library / Framework | Telemetry | Active SDK Interception |
|---|---|---|
| OpenAI | Yes | Yes |
| Anthropic | Yes | Yes |
| LangChain | Yes | Yes |
| CrewAI | Yes | Yes |
| LiteLLM | Yes | Yes |
| LangGraph | Yes | Yes |
Google Gemini (google-generativeai) | Yes | No |
| Cohere | Yes | No |
Gemini and Cohere telemetry is captured automatically by Maev's observability layer, but the SDK does not yet actively intercept or redirect their API calls. Full interception support is planned.
See Running Agents for working examples.
Current SDK interception surfaces:
- OpenAI:
chat.completions.create(...),responses.create(...), and all Azure variants (AzureOpenAI,AsyncAzureOpenAI) - Anthropic:
messages.create(...) - LangChain:
BaseChatModel.invoke(...)andBaseChatModel.ainvoke(...)on all subclasses - CrewAI: detected via
.kickoff()method on Crew objects - LiteLLM:
completion(...),acompletion(...)via callback hooks;api_baseis redirected to the Gateway whengateway=True
Telemetry support
Maev's observability layer captures data from many additional providers and frameworks out of the box:
| Library / Framework | Telemetry |
|---|---|
| Google Gemini | Yes |
| AWS Bedrock | Yes |
| Cohere | Yes |
| Mistral | Yes |
| Groq | Yes |
| Together AI | Yes |
| Ollama | Yes |
| LlamaIndex | Yes |
| Haystack | Yes |
| AutoGen | Yes |
| LiteLLM | Yes |
For these integrations, Maev can still capture spans, prompts, completions, tokens, cost, latency, and errors. Active SDK interception (retries, loop detection, gateway routing) depends on whether the library is in the patched targets list.
What gets captured per library
For every LLM call, Maev captures:
- Model name and version
- Input prompt (truncated at 4000 characters per event)
- Output completion (truncated at 4000 characters per event)
- Token counts (prompt, completion, total)
- Cost (computed from token counts and model pricing)
- Latency in milliseconds
- Any error messages
For tool/function calls, Maev additionally captures:
- Tool name
- Tool input arguments
- Tool output or error
Not seeing your library?
Maev's observability layer works with any library that follows standard tracing conventions, which means many more providers work automatically. If your library is not in the list above, open an issue or contact support.
All captured data is processed server-side. Nothing is stored locally on your machine.