Skip to main content
object
OpenTelemetry export over OTLP/HTTP.
string
The OTLP/HTTP endpoint. Telemetry is off without one.
enum<string>
default:"http/protobuf"
Wire format.Available options: http/protobuf, http/json
integer
Export timeout in seconds.
string
The service.name resource attribute.
map<string, string>
Headers sent with every export, such as auth.
map<string, string>
Extra resource attributes.
boolean
default:"true"
Include prompts and responses in spans.
boolean
default:"false"
Export logs as well as traces.
Go to full reference
string
The OTLP/HTTP endpoint to export to.Overrides telemetry.endpoint in config.toml.
enum<string>
default:"http/protobuf"
The export wire format.Overrides telemetry.protocol in config.toml.
integer
Export timeout, in milliseconds.Overrides telemetry.timeout_secs in config.toml.
string
Extra headers as comma-separated key=value pairs, added to the configured ones.Overrides telemetry.headers in config.toml.
string
The service.name resource attribute.Overrides telemetry.service_name in config.toml.
Go to full reference

Send traces to Langfuse

Point endpoint at your Langfuse region’s OTLP URL and authenticate with your project’s API keys in headers:
  • endpoint is the base URL. graph adds /v1/traces itself. For the US region, use https://us.cloud.langfuse.com/api/public/otel.
  • Set LANGFUSE_AUTH in your environment to the base64 encoding of <public key>:<secret key>. Header values can reference environment variables, so the keys stay out of the file.
  • x-langfuse-ingestion-version = "4" makes traces show up in Langfuse right away instead of after a delay.
Langfuse groups the traces of one thread, or one workbench session, into a session.

What a trace contains

  • One trace per run or turn; steps, tool calls and model calls as spans
  • Model, token usage and cost on each model call
  • capture_content controls whether prompts and responses are included; logs exports logs too

Check it worked

  • Run graph plan run or graph ask and find the trace in the backend