Provon Integrations

Provon integrates through protocols first, then through workflow destinations. Agent code can keep using familiar model and telemetry clients, while review data, findings, and regression examples flow into the systems your team already uses.

Integration Model

flowchart LR
  agent["Agent app"] --> gateway["OpenAI-compatible Gateway"]
  agent --> otlp["OpenTelemetry / OTLP"]
  gateway --> evidence["Production evidence"]
  otlp --> evidence
  evidence --> workflows["Provon workflows"]
  workflows --> notify["Notifications"]
  workflows --> data["Data destinations"]
  workflows --> export["Blob export"]

Protocol Integrations

OpenAI-Compatible Gateway

Use Provon as the model-call surface for agent applications that already speak the OpenAI API shape. The Gateway adds routing, usage policies, guardrails, trace capture, and provider key management without requiring every agent to learn a new SDK.

Typical uses:

• Route a model call through configured provider keys.
• Compare cloud, local/private model provider, and supply-node targets behind one contract.
• Capture Gateway attempts, errors, token usage, and cost into trace evidence.
• Apply fallback, retry, circuit breaker, and load-balancing policies.

OpenTelemetry / OTLP

Send traces, logs, and metrics through OTLP so Provon can build trace rollups, waterfalls, conversation views, user views, and workflow triggers from production evidence. See opentelemetry.md for language examples, collector notes, and troubleshooting.

export PROVON_API_KEY=your_project_api_key
export OTEL_EXPORTER_OTLP_ENDPOINT=http://127.0.0.1:3000
export OTEL_EXPORTER_OTLP_HEADERS="Authorization=Bearer $PROVON_API_KEY"
export OTEL_EXPORTER_OTLP_PROTOCOL=http/protobuf

Workflow Destinations

Workflow destinations are where Provon turns observations into action.

Notifications

Use notification nodes when a trace, eval, or usage policy needs human attention.

Data Destinations

Use data write nodes when findings should become durable review records.

Supported workflow records include findings, dataset examples, audience profiles, and simulation study results.

Blob Export

Blob-storage integrations are intended for archival, analysis, and downstream data pipelines. Current configuration supports S3, Azure Blob, Google Cloud Storage, and custom object storage.

Access And Secrets

Provon keeps runtime access explicit:

• Project API keys control telemetry read/ingest, Gateway invocation, model pricing access, workspace

read access, and model provider management.

• Provider-scoped API keys support supply-side operations such as manifest submission.
• Per-project provider credentials and integration secrets are stored as encrypted scoped secrets.
• Workspace sign-in can be backed by GitHub or Google.

Common Patterns

Trace To Review

1. Capture a production trace through Gateway or OTLP.
2. Match the trace with a workflow trigger.
3. Run a rule or LLM evaluator.
4. Write failed cases to Notion, Lark Base, monday.com, or Airtable.
5. Notify the owning team in Slack, Teams, Lark, or email.

Cost Alerting

1. Define Gateway usage policies for requests, tokens, or estimated cost.
2. Let Gateway reserve budget before each request and settle actual usage afterward.
3. Trigger workflows on threshold or violation events.
4. Notify owners or write the event to an operations table.

Failure To Dataset

1. Start from a failure trace or finding.
2. Extract a dataset example from the trace and root span.
3. Write the example into an external review database.
4. Use it as a regression asset during prompt, model, or routing changes.

Current Scope

Provon focuses on production evidence, review workflows, notifications, data write-back, and model traffic control. It does not currently provide a general-purpose enterprise connector framework, MCP registry, built-in RAG pipeline, or turnkey ERP/CRM sync layer.