Skip to content

Run Your Own Agent or Backend

Purpose

Run code in your own process using stock model/MCP clients and OTLP. External code does not need a Builder Agent, publication, managed execution context or Agent ID.

Prerequisites

An ordinary user/service-account virtual key in the intended Team. Its operations/resources must allow the chosen model/tool and telemetry.ingest for exported spans. Use a service account for durable backend ownership. Obtain the inference and Collector endpoints separately.

Procedure

  1. Create/review a virtual key and store its one-time secret in your process's secret configuration.
  2. Configure an OpenAI SDK or MCP client, with no direct provider fallback. Set a finite model output bound under finite spending caps.
  3. If using PydanticAI, create MissionControlConfig(api_key=...) and an ExternalObservationContext for each observation. Follow the complete adapter guide.
  4. Propagate W3C context and export completed spans through OpenTelemetry. Standalone OTLP requires no preceding inference.
  5. Inspect the actual Team's Requests/Traces/Usage. Session and customer labels use the key's server-resolved namespaces, not uploaded authority fields.

Expected result

Your application remains external; Mission Control governs physical model/tool work and authenticates telemetry independently. Rotating a secret retains its logical key/producer and accounting identity. A service account survives its creator's departure.

Verification

Make one allowed bounded call, one disallowed resource call and a standalone export. Check canonical Request evidence, a permitted Team Trace and a forbidden foreign lookup. A successful application result does not prove telemetry delivery.

Failure recovery

Reload the process after credential or configuration changes. Correct key/member/service/Team/policy denial rather than registering an Agent as a workaround. Retry failed telemetry independently; do not repeat paid work merely to obtain spans.