Skip to content

Get started with Provon

Connect your agent, inspect what it did, and review supported failures with the evidence attached. Use the AI Gateway when you also want routing and guardrails on model calls.

View as Markdown Open the plain-text version of this page.

Choose a path#

Your starting point Use What changes
You want routing, limits, or model-call guardrails AI Gateway Point your supported client at Provon.
You already instrument your application OpenTelemetry Send OTLP telemetry to your project.
You have Claude Code, Codex, or Pi transcripts Transcript sync Upload supported persisted evidence with the CLI.
You need agents to call tools through Provon MCP Gateway Publish tools and authorize the caller.

Compare capture paths before changing an existing integration.

How the workflow fits together#

Your agent Capture traces Inspect behavior Review a Finding Repair in your workflow Check new evidence
Step What you do Check before continuing
Connect Select a project and configure a capture path. The request or uploaded evidence appears in Traces.
Inspect Review messages, spans, usage, and errors. The trace includes the goal and activity relevant to your question.
Diagnose Enable diagnostic Rules or analyze existing evidence. Review any Finding against its cited evidence; none may be needed.
Repair Hand a supported Finding to your issue tracker. The owner has a scoped problem and verification context.
Verify Run the agent again after the change. New traces show whether the original behavior improved or regressed.

What Provon controls#

Capability Scope
AI Gateway Routes model calls, applies configured limits and guardrails, and records requests when tracing is enabled.
Tracing Collects and displays evidence from supported Gateway, OTLP, and transcript capture paths.
Findings Presents supported diagnostic problems for review, confirmation, dismissal, and repair handoff.
Playground Compares model selections, prompts, messages, and generation parameters through the AI Gateway.
MCP Gateway Governs calls to tools published through its endpoint; it does not control unrelated tool execution.

Guardrails only apply at the boundary you connect. OTLP and transcript uploads do not intercept model calls, browser activity, or tool execution. Gateway traces alone may not include tool execution outside the model-call path; add application instrumentation when that evidence matters.

Provon does not edit your repository, apply repairs, or treat a generated diagnosis as proof. Your team reviews Findings and owns changes in its existing development workflow.

Continue#