Frameworks
LangChain and LangGraphTrace LangChain chains and LangGraph graphs with OpenInference: each invocation becomes a run with its model calls, token counts, tool calls and graph steps.CrewAITrace a crew's agents, tasks and tools with OpenInference, and its model calls and token counts with the instrumentation for its model provider.OpenAI Agents SDKGive the Agents SDK an instrumented OpenAI client: every model call is recorded with its token counts, and checkpoints and approvals work inside your tools.Claude Agent SDKRecord each Claude Agent SDK session as a run with its reported token usage and cost, log tool calls, and require a person's approval before sensitive tools run.Vercel AI SDKExport the AI SDK's OpenTelemetry spans to Toolcaise: every generation step is a model call with token counts, and every tool call is recorded by name.
Choosing a path
There are three ways in. They differ in what they can see and what they can do.
| Path | Sees | Can decide | Recommended for |
|---|---|---|---|
| OpenTelemetry | Runs, model calls with token counts, tools and steps, from the framework's own instrumentation | Nothing: it only reports | LangChain and LangGraph, CrewAI, Vercel AI SDK |
| Python or Node SDK | What you wrap: runs, spans and outcomes, and model calls through an instrumented OpenAI or Anthropic client | Pause, resume and cancel at checkpoints, and approval of exact actions | OpenAI Agents SDK, Claude Agent SDK |
| Toolcaise Connect | Every model call made through OpenAI- or Anthropic-compatible APIs, measured by the connector | Stops the run at time, token, cost or call limits; pause, resume or stop from the dashboard | Any agent you start from a command line |
They combine. An OpenTelemetry trace and an SDK run are separate runs in Toolcaise, so use the SDK where a person has to decide and OpenTelemetry for the full picture.
Any OpenTelemetry exporter
Any agent that exports OpenTelemetry traces over OTLP/HTTP can report to Toolcaise with environment variables alone:
Environment
OTEL_EXPORTER_OTLP_TRACES_ENDPOINT=https://toolcaise.com/api/v1/otlp/v1/traces
OTEL_EXPORTER_OTLP_HEADERS=Authorization=Bearer%20<agent credential>
OTEL_EXPORTER_OTLP_COMPRESSION=gzip
OTEL_METRICS_EXPORTER=none
OTEL_LOGS_EXPORTER=noneThe %20 is required: header values in these variables are URL-encoded. Toolcaise reads the OpenTelemetry GenAI conventions, OpenInference, OpenLLMetry and the Vercel AI SDK's attributes; a trace with none of them is accepted and not stored.