# AI monitoring (Insights)

> https://docs.errora.org/en/products/insights

Insights is LLM observability built on the tracing pipeline. The official Sentry SDKs'
**AI-agents** and **MCP** integrations emit OpenTelemetry `gen_ai.*` and `mcp.*` spans;
Errora projects those spans into a queryable table and gives you two dashboards.

## Setup

There is no separate product to enable — turn on tracing and the AI integration for your
stack.

```python
import sentry_sdk

sentry_sdk.init(
    dsn="…",
    traces_sample_rate=1.0,       # AI traffic is low-volume and high-value
    send_default_pii=False,       # set True only if you want prompts recorded
)
```

The Python SDK instruments OpenAI, Anthropic, LangChain and the Agents SDKs automatically
once tracing is on. In JavaScript, `Sentry.init({ tracesSampleRate: 1 })` plus the
`openAIIntegration()` / `anthropicAIIntegration()` / `vercelAIIntegration()` integration does
the same.

For MCP servers, the SDK's MCP integration emits `mcp.*` spans per JSON-RPC method — the
same spans Errora's own [MCP server](/en/products/mcp) produces.

:::caution Prompts and PII
Prompt and completion text is only recorded when the SDK is configured to send it
(`send_default_pii=True` or the integration's explicit option). Leave it off unless you have
a reason — prompts routinely contain user data.
:::

## Manual spans

If you call a model directly, wrap it yourself:

```python
with sentry_sdk.start_span(op="gen_ai.chat", name="summarize") as span:
    span.set_data("gen_ai.system", "openai")
    span.set_data("gen_ai.request.model", "gpt-4o-mini")
    response = client.chat.completions.create(...)
    span.set_data("gen_ai.usage.input_tokens", response.usage.prompt_tokens)
    span.set_data("gen_ai.usage.output_tokens", response.usage.completion_tokens)
```

## The AI Agents dashboard

- **Agent runs** over time, with duration
- **LLM calls** — count and latency
- **Tokens by model**, split cached / not cached
- **Tool calls** — which tools your agent actually reaches for
- **Recent runs**, each opening a **per-run span timeline**: the agent's steps, the model
  calls it made and the tools it invoked, in order

## The MCP Server dashboard

- **Traffic** over time
- **By client** — which agent is calling you
- **By method and transport**
- **Most-used tools, resources and prompts**

## API

```text
GET /api/v1/projects/{project_id}/insights/agents
GET /api/v1/projects/{project_id}/insights/agents/runs
GET /api/v1/projects/{project_id}/insights/agents/runs/{trace_id}
GET /api/v1/projects/{project_id}/insights/mcp
```

## Troubleshooting

| Symptom | Cause / fix |
| --- | --- |
| Dashboards are empty | Tracing is off, or the SDK's AI integration isn't active — check that transactions contain `gen_ai.*` spans |
| Token counts missing | The provider response didn't carry usage, or a manual span didn't set `gen_ai.usage.*` |
| Runs look truncated | The trace was sampled away mid-run; raise `traces_sample_rate` for AI workloads |
| Prompts visible when they shouldn't be | An integration option is sending PII — audit `send_default_pii` and per-integration flags |