> ## Documentation Index
> Fetch the complete documentation index at: https://glassray.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# Quickstart

> Connect your traces and your code, and let Glassray find recurring deviations and propose fixes - end to end.

Glassray runs a continuous loop over your agent's production traffic: it learns what your agents are *supposed* to do, scans your traces for recurring misbehavior, and proposes code-level fixes. Here's how to get it running.

<Note>
  Glassray is in **private alpha**. [Join the alpha](https://glassray.ai/waitlist) and we'll connect your traces, walk through your agent's code, and get the loop running on your production data.
</Note>

<Steps>
  <Step title="Connect a trace source" icon="plug">
    Point Glassray at the traces your agents already emit - no instrumentation rewrite. Connect [Langfuse](/langfuse), [LangSmith](/langsmith), or [PostHog](/posthog) (Glassray pulls on a cadence), or push [OpenTelemetry](/otlp-ingestion). See [Traces](/traces) for the concept.
  </Step>

  <Step title="Connect your code" icon="github">
    Glassray reads your code to learn intended behavior. [Connect GitHub](/github) and it generates a [spec](/spec) from your repo - or, if you'd rather not connect, [bring your own spec](/spec#bring-your-own-spec).
  </Step>

  <Step title="Glassray learns your system" icon="brain">
    Glassray sorts your traces into [flows](/flows) - groups of traces that run the same agent chain - and uses the spec as the **rubric** for what each flow should do.

    <Info>
      Every deployment is different. Two teams running the "same" framework with different prompts and rules are effectively running different programs, with different failure modes. Glassray learns yours.
    </Info>
  </Step>

  <Step title="Review deviations" icon="ghost">
    Glassray scans your traces against the spec and clusters what it finds into recurring [deviations](/deviations) - *types* of misbehavior that show up across many runs, not one-off errors. The ones worth catching: **silent failures** (the output looks right, but the process doesn't hold up) and **intent mismatches** (a step did something the code says it shouldn't). You can also surface deviations from [Slack](/slack).
  </Step>

  <Step title="Accept a fix" icon="code-pull-request">
    Accept a deviation and Glassray proposes a **code-level fix** as a diff with the failing trace attached - review it like any pull request - then runs a **deep search** across your history to find every other place it shows up.

    <Check>
      Every merged fix tightens the loop - your agents get better at avoiding that deviation, and Glassray keeps scanning.
    </Check>
  </Step>
</Steps>

## Next steps

<CardGroup cols={2}>
  <Card title="Flows" icon="diagram-project" href="/flows">
    How Glassray groups your traces into flows - the unit it tracks quality on.
  </Card>

  <Card title="Ask Glassray" icon="comments" href="/ask-glassray">
    Explore your traces in natural language.
  </Card>

  <Card title="Connect over MCP" icon="plug" href="/mcp-server">
    Query your traces, flows, and deviations from your coding agent.
  </Card>

  <Card title="Deviations" icon="ghost" href="/deviations">
    What a deviation is and its review lifecycle.
  </Card>
</CardGroup>