Nyraxis AI

AI Agent Integration Guides

Choose an integration by the operation you need to evaluate or record. Direct evaluation, automatic input interception, output release, and tracing have different boundaries. A successful import or a recorded span does not establish that an operation was enforced.

Start with one direct evaluation

The quickstart walks through an ingest key, a configured policy, synthetic text, and the returned decision. Review completion, errors, and required policy execution before allowing the next model or tool operation.

  • Python SDK: client configuration, direct evaluation, supported wrappers, and lifecycle.
  • Node.js SDK: direct evaluation and supported OpenAI method interception.
  • HTTP evaluation API: request fields, authentication, decisions, and diagnostics.

Choose a framework or provider recipe

Recipes show a specific application boundary. Verify the installed package version and supported operation in the integration support matrix.

  • LangGraph: a model and read-only tool journey with explicit input and output gates.
  • LangChain: evaluate requests around the application-owned invocation.
  • CrewAI: gate the workflow input and inspect its buffered result.
  • LlamaIndex: put evaluation around a selected retrieval workflow.
  • Vercel AI SDK: server-side checks around buffered generation.
  • All integration recipes: provider and server-framework examples.

Keep evaluation, enforcement, and tracing separate

An evaluation describes the configured checks. An application gate decides whether a model call, tool action, or response release should proceed. Tracing records the spans your instrumentation sends; uninstrumented operations remain outside that record.

Read the enforcement guide for supported interception and failure behavior, and the tracing guide for span capture and lifecycle. Streaming output needs an explicit release strategy; a buffered-output recipe does not establish a streaming guarantee.

Verify a benign, unsafe, and unavailable-check case

  1. Run a synthetic request that should complete and continue. Confirm the actual model or tool operation occurs.
  2. Run a known synthetic failure. Confirm the application stops at the intended boundary and preserves useful evidence.
  3. Exercise an unavailable evaluator or detector. Confirm the chosen failure behavior and distinguish incomplete evaluation from a safe result.
  4. Save a regression assertion and retest it after relevant changes.

Review the data boundary before using customer content

Evaluation sends text to the configured API; selected checks may use additional model endpoints. Capture and retention settings govern stored content. Review data handling and private deployment operations before a broader rollout.