Wherever your agents run. However they're built.
Trefur is built on OpenTelemetry GenAI Semantic Conventions — instrument once and capture everything. Drop-in adapters for the frameworks you use, auto-instrumentation for every major model vendor, OAuth connectors for the SaaS systems your agents touch.
The same OAuth connectors inventory the AI you didn't deploy — copilots and third-party agents already operating in your tools. See governance & safe AI.
Pick the category closest to where you are.
Drop-in instrumentation for the framework you're already using.
Add the SDK, set your key, ship. The framework's existing callbacks emit OpenTelemetry-canonical spans for every chain step, tool call, retry, and hand-off. No re-platforming.
What you see
- Every chain or graph step, with the prompt the model saw and the response it produced
- Every tool call, with arguments, status, latency, and result
- Every retry, with the reason and the backoff
- Every sub-agent spawn and the data that crossed the boundary
- Every error and where in the graph it surfaced
What you get out of it
- Debug a failing run in seconds instead of digging through four log streams
- Catch silent regressions when a model swap breaks a tool-calling shape
- Stay on the framework you picked — no vendor lock-in to a tracing layer
LangChain
Trace every chain step, tool call, retry, and agent decision. LangChain callbacks emit OTel-canonical spans automatically.
CrewAI
See the crew. See the tasks. See the hand-offs. Crews and their delegations show up as a connected trace, not isolated LLM calls.
AutoGen
Multi-agent conversations become navigable. Group chats, nested agents, and tool invocations land in a single trace with role attribution.
Pydantic AI
Typed function calls and structured outputs map cleanly to OTel GenAI spans. Validation failures surface as first-class events.
MCP
Servers, clients, tool calls, and resource fetches — every step in an MCP interaction is captured. Works for stdio, SSE, and HTTP transports.
One trace shape across every model vendor your agents call.
Auto-instrument the LLM-call layer the moment you initialize the SDK. Tokens, latency, cost, and prompt/response context attached to the step that emitted them — across every provider in the same canonical shape.
What you see
- Tokens in, tokens out, cost in dollars per call
- Model name and version, decoded from the actual response
- Time-to-first-token and total latency
- Prompt + completion (redactable, off by default in the collector if you want it that way)
- Tool-call shape across the provider's native function-calling API
What you get out of it
- Compare vendor cost and latency on real production traffic, not on a benchmark
- Move workloads between providers without breaking your dashboards or alerts
- Spot a model regression the day the provider ships a silent update
OpenAI
GPT-4, GPT-5, function calls, structured outputs, the Assistants API. Auto-instrumentation captures the call and the cost.
Anthropic
Claude family, tool use, computer use, prompt caching. Every call traced with its cache hits, tool decisions, and final output.
Gemini
Google's Gemini models — captured through the Google AI SDK or Vertex AI. Streaming, tool use, multimodal calls all flow into the trace.
Groq, MiniMax, others
Any OpenAI-compatible API works out of the box. Custom providers attach via OpenTelemetry GenAI semantic conventions.
Observe agents on the cloud platform your security team already approved.
Bedrock, Azure OpenAI, Vertex, Cortex, Mosaic. Your cloud picks the model and handles the IAM and residency for the model layer. Trefur captures what your agents do on top of it — hosted in Singapore by default, with regional data residency available on request.
What you see
- Model invocations across the cloud's managed AI services
- Tool-use payloads and Knowledge Base / RAG retrieval calls
- Cross-account, cross-region, and cross-project usage attributed to the run that triggered it
- Cost broken out by deployment, environment, and team
What you get out of it
- Keep your cloud's IAM, VPC, and model-layer residency posture — Trefur hosts what it observes in Singapore, with regional residency on request
- Compare a managed-cloud workload against a direct-vendor workload on the same trace shape
- Surface cross-account spikes before they show up on next month's invoice
AWS Bedrock
Claude, Llama, Mistral, Titan, Nova on Bedrock. Auto-trace InvokeModel + Agents + Knowledge Bases + Guardrails activity.
Azure OpenAI
OpenAI models deployed in Azure. Captures per-deployment usage, content filters, and Assistants activity, all inside your tenant.
Google Vertex AI
Vertex AI and the Gemini API on GCP — agent calls, tool use, and grounded responses captured through the Vertex SDK.
Databricks Mosaic AI
Mosaic AI Agents and the MLflow tracing surface — agent activity in your lakehouse, traced into Trefur.
Snowflake Cortex
Cortex AI and Snowpark agent activity inside your Snowflake account, surfaced in Trefur alongside everything else.
See agents touching real business systems — files, messages, tickets, deals.
Connect via OAuth. Trefur pulls activity from the platform's admin APIs and stitches it into the agent trace. No SDK required in the agent itself — useful for AI assistants and copilots embedded inside the SaaS.
What you see
- Which agents are reading or writing in each platform, and on whose behalf
- Bot, integration, and AI-assistant activity from the platform's audit log
- Third-party OAuth grants and the agents they enable
- Activity volume per team, per workspace, per environment
What you get out of it
- Inventory every agent and copilot operating inside your tools — including ones you didn't know about
- Tie a customer interaction in support / sales / finance back to the agent that handled it
- Catch quiet failures — agents stuck mid-workflow that nobody noticed
Google Workspace
Audit log, Gemini activity, third-party OAuth-grant inventory. See AI inside Gmail, Drive, Docs, Meet.
Microsoft 365
Audit logs and Copilot prompt counts across Teams, Outlook, SharePoint, OneDrive, Word, Excel, PowerPoint.
Slack
Bots and integrations active in your workspace — what they ran, where they posted, who triggered them.
GitHub
Coding agents, Copilot, automation bots. PRs they opened, files they touched, checks they ran.
Salesforce
Einstein, Agentforce, Flow, and AI-driven automation across your CRM.
HubSpot
Workflow and AI-assistant activity inside HubSpot's marketing and sales hubs.
Zendesk
Support-agent and AI-macro activity across tickets, side-conversations, and the help-center bot.
Intercom
Fin AI and Custom Bot activity inside your Intercom workspace.
Atlassian
Jira and Confluence — Atlassian Intelligence, installed-app activity, automation runs.
Built on open standards. No lock-in, ever.
Trefur reads OpenTelemetry GenAI Semantic Conventions natively. If your framework, runtime, or in-house code emits OTel, Trefur reads it. Add Trefur as a second exporter and keep your existing backend intact.
What you see
- Standard OTLP/HTTP and OTLP/gRPC at /v1/traces, /v1/metrics, /v1/logs
- GenAI semantic-convention spans from any compliant emitter
- Spans from custom or in-house frameworks the moment they emit OTel
What you get out of it
- Adopt without forcing a re-platform — your existing observability pipeline stays in place
- Fan out to Trefur alongside Datadog, Honeycomb, Grafana, or any other OTLP-compatible backend
- Future-proof against the next framework — if it emits OTel, you're already covered
OpenTelemetry GenAI
OTLP/HTTP and gRPC endpoints, GenAI semantic conventions, drop-in for any compliant emitter.
SDK languages
Python, JavaScript, Go, Rust, Java, .NET — drop-in adapter SDKs in every major language.
Self-hostable collector
Single Go binary inside your network. Local batching, redaction, on-disk buffering, fan-out.
Custom or in-house framework?
If you emit OpenTelemetry traces with GenAI semantic conventions, Trefur reads them. No vendor lock-in. No proprietary format. The framework you build next year is already supported.
Read the docsStart free. First trace in five minutes.
Drop the SDK in, set your key, run your agent. The trace appears.