Integrations
Evals.si does not replace your agent stack. It evaluates it. Every integration uses one of five doors, so a platform that supports any of them works without custom code.
| Door | Direction | Use it to |
|---|---|---|
| OTLP traces | your platform → Evals.si | Score live traffic with an online policy and promote failures into regression datasets; LangGraph, CrewAI, the OpenAI Agents SDK, LlamaIndex and the Claude Agent SDK are tested |
| Trace sources | Evals.si → your trace store | Read the traces you already keep in MLflow, Phoenix or Langfuse, score them, and write the scores back to the same traces (guide) |
| Agent targets | Evals.si → your agent | Run your agent on sandboxed tasks and benchmarks over A2A, MCP, an OpenAI Responses API, plain HTTP or a CLI |
| Inline guardrails | your gateway → Evals.si | Redact, evaluate and block content while a request is in flight |
| MCP and the API | your tools → Evals.si | Let coding agents, CI and dashboards start runs and read results over gRPC, HTTP or MCP |
Guides
Section titled “Guides”- agentgateway: trace export and inline guardrails for LLM, MCP and A2A traffic.
- Kubernetes: install with Helm, drive runs and policies with CRDs, and wire in-cluster agents.
- Agent platforms and studios: connect a visual or low-code agent builder, or any agent framework.
- Agent frameworks: the instrumentation to use for each tested framework, and what it records.
- Trace sources: score traces kept in MLflow, Phoenix or Langfuse.
- Integrate an agent harness: install beside an agent harness, score every workflow online, and gate deploys.
For the plugin side, see Plugins and Wasm evaluators.