Integrations

Works with what you already run.

Two settings with OpenTelemetry, or a few lines of SDK for the most detail. Switch Runback off below and see what you'd lose.

.envRunback on

 # Any framework that emits OpenTelemetry GenAI spans: Python, Go, Java, Node
+export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT="https://runback.dev/api/otel/v1/traces"
+export OTEL_EXPORTER_OTLP_TRACES_HEADERS="authorization=Bearer <RUNBACK_API_KEY>"
  
 opentelemetry-instrument python your_agent.py

research-email-agent · run failedclick any step ↓

Tool call · send_email — failed here

tool: send_email
status: error

input

{ "to": "alex[at]acme.co", "body": "Next.js 16 summary…" }

Invalid recipient — the model wrote the address in plain text ("[at]") instead of a valid email. The run failed here.

Illustrative run. On this path a model step can be re-issued, not replayed.

What OpenTelemetry records, and what it can't do

Code changesNoneYour framework's own instrumentation sends the spans; no Runback package.
ContextWhat's exportedMessages, system prompt and tool definitions when the instrumentor exports them (GenAI, OpenLLMetry, OpenInference).
RedactionOn arrivalRedacted by Runback when the spans arrive, not inside your process.
Block before it runsNoSpans arrive after the call ran, so nothing can stop it.
ReplayRe-issue onlyA model step can be re-sent to a live model. The whole run can't be re-run with the same tool results: they aren't kept.
ArrivesAbout a minuteA run is written once its trace goes quiet, about a minute after the root span.
Signed recordYesEach run is hash-chained and can be exported as a signed record.

Start free →Read the OpenTelemetry docs →

Any framework

Point OpenTelemetry at it.

Two environment variables, then your framework's own instrumentation. No Runback package.

OpenAI
Anthropic
Gemini
Llama
Mistral AI
Vercel AI SDK
LangChain
CrewAI
Ollama
OpenTelemetry
Python
TypeScript

What Runback plugs into. Not customers: we have none to name yet.

Protobuf or JSON, default settings. An OTel run's steps can be re-issued, not replayed exactly.¹

Two environment variablesShell

export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT="https://runback.dev/api/otel/v1/traces"
export OTEL_EXPORTER_OTLP_TRACES_HEADERS="authorization=Bearer <RUNBACK_API_KEY>"

CrewAI

OpenTelemetry · tested recipe

OpenInference's CrewAI and OpenAI instrumentors together; crewai 1.10.1 or later, below which the instrumentor silently records nothing. Model steps arrive with messages, system prompt, tool definitions, response text, tool calls and tokens, and tool steps in full.

CrewAIPython

pip install opentelemetry-sdk opentelemetry-exporter-otlp-proto-http

from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
tp = TracerProvider()
tp.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))   # protobuf, the default
CrewAIInstrumentor().instrument(tracer_provider=tp)
OpenAIInstrumentor().instrument(tracer_provider=tp)

LlamaIndex

OpenTelemetry · tested recipe

OpenInference's LlamaIndex instrumentor. Pass include_usage in stream_options to the OpenAI LLM or tokens are zero. Each model call currently arrives as two model steps: one complete, with tokens and tool calls, and a duplicate with no tokens whose response text is the instrumentor's repr of the messages.

LlamaIndexPython

pip install opentelemetry-sdk opentelemetry-exporter-otlp-proto-http

from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
tp = TracerProvider()
tp.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))   # protobuf, the default
LlamaIndexInstrumentor().instrument(tracer_provider=tp)
llm = OpenAI(model=..., additional_kwargs={"stream_options": {"include_usage": True}})

Pydantic AI

OpenTelemetry · tested recipe

Pydantic AI's own instrumentation. Model steps arrive with messages, system prompt, tool definitions, response text, tool calls and tokens, and tool steps in full.

Pydantic AIPython

pip install opentelemetry-sdk opentelemetry-exporter-otlp-proto-http

from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
tp = TracerProvider()
tp.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))   # protobuf, the default
Agent.instrument_all(InstrumentationSettings(tracer_provider=tp))

Google ADK

OpenTelemetry · tested recipe

OpenInference's ADK instrumentor; google-adk 2.10 or later. Without it, ADK's native spans make tool calls show up as model calls. Model steps carry their tool calls; the provider shows as gcp.vertex.agent, not the model's vendor.

Google ADKPython

pip install opentelemetry-sdk opentelemetry-exporter-otlp-proto-http

from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter
tp = TracerProvider()
tp.add_span_processor(BatchSpanProcessor(OTLPSpanExporter()))   # protobuf, the default
GoogleADKInstrumentor().instrument(tracer_provider=tp)

Anything else (raw OpenTelemetry)

OpenTelemetry · any language

If it exports GenAI spans over OTLP/HTTP, protobuf or JSON, Runback reads it. Set the two environment variables; the exporter's defaults are fine.

Anything else (raw OpenTelemetry)Shell

export OTEL_EXPORTER_OTLP_TRACES_ENDPOINT="https://runback.dev/api/otel/v1/traces"
export OTEL_EXPORTER_OTLP_TRACES_HEADERS="authorization=Bearer <RUNBACK_API_KEY>"

If it exports OpenTelemetry GenAI spans over OTLP/HTTP, it works. Tell us what you use and we'll confirm the mapping — read the guide or open a sample run.

Most detail

Or add a few lines of SDK.

Full context capture and in-process redaction. In TypeScript, policy blocks a tool call before it runs and the whole run can be replayed (Enterprise).

Pick your SDK in the setup above ↑Read the full walkthrough →Already on Traceloop? See how it compares →

Any model

Record any model.

Whatever model you call is recorded. Replay runs on OpenAI, Anthropic and Groq-hosted open models today.

Connects out

It lives in your stack, not another console.

The record goes to the systems your organisation already runs.

IdentityOkta · Azure AD / Entra · Google Workspace · Auth0
native
CI / CDGitHub Actions
native · GitHub
Incident & chatSlack · PagerDuty · any webhook
alerts · webhooks
Regulatory evidenceEAAPL — AU Enterprise AI Patterns
read-only API
SIEMSplunk (HEC) · Microsoft Sentinel · any webhook
native · webhooks
Data exportREST API · JSON
API
Your cloudAWS · Azure · GCP · on-prem
self-host

Don't see yours? Everything is reachable through the open REST API and outbound webhooks — tell us what you run and we'll confirm the wiring. AU regulatory evidence pack? See the EAAPL integration →