Build self-improving loops for your agents.

/dev/provon

Route model calls through one Gateway, capture Gateway and OTLP evidence, and turn risky traces into loop feedback.

Star Provon on GitHub
Open-source core. Self-host on Node, desktop, or Workers. read the docs
Provon / Get Started
Trace evidence
1,204
Risk signals
18
Loop artifacts
43

Risk Signals

Trace match · Evaluator · Artifact
Trace IDSignalPolicyState
tr_8f9a2bRepeated tool callsGuardrailBlocked
tr_1c4d5eHigh token usageCost policyLoop fired
tr_3b6a9cProvider fallback usedGatewayCaptured
tr_7e2f1dSuspicious final answerEvaluatorFinding
Use the protocols agents already speak. Deploy where the traffic lives.
OpenAI-compatible Gateway
OpenTelemetry / OTLP
Node, Desktop, or Worker
MIT licensed core

Get Started

Start with the AI Gateway. Keep OpenTelemetry when you already have it.

Provon captures evidence in two ways. AI Gateway is the recommended path; OTLP is the compatible path for agents already instrumented with OpenTelemetry.

Recommended path. Point an OpenAI-compatible SDK at Provon to capture model calls with routing, policies, usage, and provider attempts.

Read Gateway guide
AI Gateway
OpenAI-compatible model call
Recommended
POST /gateway/v1/chat/completions
model: openai/gpt-5-mini
Authorization: Bearer project key
Endpoint
/gateway/v1
Contract
OpenAI-compatible
Capture
automatic
Policy
cost + guardrail
Captured prompt, response, provider attempt, tokens, cost, latency, and errors.
checkout-agent
OpenAI-compatible request
Connected
Gateway
/gateway/v1
OTLP
/v1/traces
Project
agent-prod
Policy
cost + guardrail
Trace coverage Gateway + OTLP
Authorization: Bearer project key

Loops

How It Works

Go from raw agent traffic to trace-backed feedback artifacts in seconds.