Agents, tools and prompts
Tools
Agents choose from the tools you select on the agent. Each tool call is checked against the agent's tool list, budgeted against max_steps, and its result is treated as untrusted data.
| Source | How to add | Approval |
|---|---|---|
| Built-in | knowledge_search (with an optional metadata filter), memory_search and others listed by GET /api/tools | Read-only |
| API actions | Register one in Build → API actions, or import many with POST /api/actions/import-openapi {spec, base_url?, operations?} | POST actions always wait for a different person |
| MCP servers | Add a server in Build → MCP servers with its url, optional key_env and a readonly flag; choose which of its tools to expose | Calls to a server that isn't readonly wait for a different person |
| Integrations | Netra read tools and Keep run_code (see integrations) | run_code always waits |
The OpenAPI importer turns GET and POST operations with flat query or JSON-body inputs into typed actions. MCP tools appear as mcp_<first 8 characters of the server id>_<tool>. MCP server hosts must be on NUVORA_PROVIDER_HOSTS.
Memory
Agents can search what the same person stored in earlier sessions with memory_search. Writing memory is a consequential step: it waits for approval like any other write. Turn on summarize_memory and the agent proposes a summary at the end of a session, which is saved only if it is approved.
Tracing
Every agent and workflow run records each step's duration. Runs shows them as a waterfall. Set NUVORA_OTEL_ENDPOINT to export the same spans over OTLP/HTTP to Jaeger, Tempo or any OpenTelemetry collector. Spans carry timings, model names, token counts and costs, never prompt or answer text.
Prompt variants and experiments
A prompt can carry variants, each with its own template and a weight. Rendering picks a variant by weight and keeps that choice stable for the same subject, so one person sees the same wording across requests.
POST /api/prompts/{id}/experiment {evaluation, variable} runs the control template and every variant over an evaluation suite's cases. The result shows each arm's score and names a winner only when a variant beats the control.