Support matrix
Every framework that sends OpenTelemetry (OTel) spans to Logfire gets the Live and Explore views, the trace waterfall, annotations, dashboards, and alerts. The specialized LLMs and Agents views need specific span attributes and parent-child relationships, so their support varies by framework.
This page summarizes what each framework’s guide describes about Logfire’s views. The determining factor is how the framework emits its spans:
- Native OTel-GenAI semconv (
gen_ai.*): the LLMs and Agents views understand these directly, so model, token, cost, and — where the framework emits them — tool and message data all populate. - Instrumentor or bridge (e.g. OpenInference / OpenLLMetry, a framework’s own OTel bridge, or a Logfire-native instrumentor): agent runs and token counts populate, but cost, the tools tab, and message/prompt content are not yet read from these conventions.
Legend: ● full · ◐ partial or framework-dependent · ○ not yet.
These frameworks emit gen_ai.* spans, so the Agents view shows runs, tokens, model, and cost; where a framework
populates only a subset, the marks below show it. Tools and message content appear when the framework emits them
(see each guide).
| Framework | Live & Explore | Agents view | Tokens & model | Cost | Tools | Messages |
|---|---|---|---|---|---|---|
| Pydantic AI | ● | ● | ● | ● | ● | ● |
| Google ADK | ● | ● | ● | ● | ◐ | ◐ |
| Strands Agents | ● | ● | ● | ● | ○ | ◐ |
| Semantic Kernel | ● | ● | ● | ● | ○ | ○ |
| AutoGen | ● | ● | ● | ● | ◐ | ◐ |
| Mastra | ● | ● | ● | ● | ○ | ● |
| Vercel AI SDK | ● | ● | ● | ● | ◐ | ● |
| Rig | ● | ● | ● | ● | ◐ | ◐ |
| Microsoft Agent Framework (.NET) | ● | ● | ● | ● | ◐ | ◐ |
| Semantic Kernel (.NET) | ● | ◐ | ● | ● | ○ | ◐ |
Detected as agents with per-run token counts. Cost, the tools tab, and message/prompt content are not yet read from these conventions.
| Framework | Live & Explore | Agents view | Tokens & model | Cost | Tools | Messages |
|---|---|---|---|---|---|---|
| Agno | ● | ● | ● | ○ | ○ | ○ |
| smolagents | ● | ● | ● | ○ | ○ | ○ |
| LangGraph | ● | ● | ● | ○ | ○ | ○ |
| OpenAI Agents SDK (Python) | ● | ● | ● | ○ | ○ | ○ |
| CrewAI 1 | ● | ● | ○ | ○ | ○ | ○ |
| Eve 2 | ● | ◐ | ● | ○ | ○ | ○ |
Detected as agent runs from a framework-specific key. Model-call metrics depend on how the framework emits its LLM calls; in the configurations we tested they are not captured as model calls.
| Framework | Live & Explore | Agents view | Tokens & model | Cost | Tools | Messages |
|---|---|---|---|---|---|---|
| Genkit (Go) | ● | ● | ○ | ○ | ○ | ○ |
| VoltAgent | ● | ● | ○ | ○ | ○ | ○ |
These emit valid OTel spans — so Live, Explore, dashboards, and alerts all work — but do not currently produce an agent-run span the Agents view recognizes (they emit an LLM or chain span, or a non-agent convention).
| Framework | Live & Explore | Agents view | Notes |
|---|---|---|---|
| Haystack | ● | ○ | Emits a chain/LLM span, not an agent-run span. |
| LlamaIndex (TS) | ● | ○ | Emits an LLM chat span only. |
| LangChain (JS) | ● | ○ | Traced via LangSmith’s own convention. |
| Instructor | ● | ○ | LLM-only; not an agent framework. |
| Eino (Go) | ● | ○ | No agent-run convention we detect today. |
| Letta | ● | ○ | Server traces flow via an OpenTelemetry Collector; no agent-run span is detected. |
| OpenAI Agents SDK (TS) | ◐ | ○ | No OpenTelemetry exporter yet; instrument the underlying model client for LLM spans only. |
| Claude Agent SDK | ● | ○ | Emits native gen_ai invoke_agent/chat/execute_tool spans (LLMs view, tokens, and cost populate), but sets no gen_ai.agent.name, so the Agents view does not detect the run. |
-
CrewAI is detected and named on the Agents view, but the CrewAI instrumentation in our tests emits no LLM spans, so token, model, and cost columns are empty. ↩
-
Eve agent runs and conversations are detected, but aggregate Agent metrics can miss turns or double-count repeated token attributes (tracking issue #28737). ↩