1. Acesse o Hub, pesquise por AI observability e selecione o app AI Observability.

   

2. Abra o app e clique em Open (ou Install, caso ainda não esteja instalado) para disponibilizá-lo na tenant.

   

1. Abra o menu rápido com Ctrl+K, pesquise por acc e selecione Access Tokens (Classic apps).

   

2. Na tela de Access tokens, clique em Generate new token.

   

3. Dê um nome ao token e selecione os scopes metrics.ingest (Ingest metrics) e openTelemetryTrace.ingest (Ingest OpenTelemetry traces). Em seguida, clique em Generate token.

   

Windows 1803+ is required for the built-in tar command used to extract the collector binary.

The extension requires outbound HTTPS (port 443) access to the following domains. All connections are outbound only — no inbound ports are opened on the developer machine.

For corporate environments: share this table with your IT/security team before installation. Missing any of these will cause silent failures (no error shown to the user — data simply stops flowing).

Required for installation (one-time):

Required for OTel Collector binary download (one-time, ~100 MB):

The exact URL pattern for the binary download is:

https://github.com/open-telemetry/opentelemetry-collector-releases/releases/download/
  v{version}/otelcol-contrib_{version}_{os}_{arch}.tar.gz

Required at runtime (continuous):

Replace * with your tenant ID, e.g.: abc12345.live.dynatrace.com

Local ports (no external access required):

Port conflicts: ports are checked before the collector starts. If 4318 or 13133 is already in use, the extension automatically finds and uses the next free port and updates the setting for you. You can still set a preferred value manually:

• dynatraceAiObs.collectorPort — controls the OTLP receiver port (default 4318)
• dynatraceAiObs.healthCheckPort — controls the health check port (default 13133)

Optional (Windows only — Python installation):

Summary for IT/security ticket:

Allow outbound HTTPS (443) to:
- github.com
- objects.githubusercontent.com
- raw.githubusercontent.com
- *.live.dynatrace.com  (replace * with tenant ID)

Optional (Windows, Python install):
- python.org
- files.pythonhosted.org

No inbound rules required.
No VPN split-tunnel changes required.

The Configurações tab contains credentials, privacy controls, local ports, and custom attributes in one place.

Panel language: a dropdown in the top-right corner of the panel lets you switch its UI between Português (BR), English, and Español at any time — the choice is saved automatically and persists across restarts (dynatraceAiObs.language). This only affects the panel itself; Command Palette entries and VS Code notifications are unaffected.

Tenant ID mode (recommended)

Enter only the tenant identifier. The generated OTLP endpoint is shown immediately below the field. The API token remains protected in the OS keychain, while the email identifies the developer in reports.

The highlighted Coleta area controls data privacy. Enabling Capturar conteúdo de prompts e respostas unlocks Dynatrace Evals; when disabled, only metadata such as model, duration, tokens, and tool calls is collected.

Full OTLP endpoint mode

Choose OTLP Endpoint completo when using a custom or non-standard endpoint. Enter the complete /api/v2/otlp URL and verify that the hostname uses .live.dynatrace.com, never .apps.dynatrace.com.

Use Validar credenciais to test the endpoint and token, or just click Salvar configurações — credentials are now validated automatically before saving, and the collector restarts on its own to apply the new values.

The Coletor tab provides day-to-day operational control without leaving the panel:

• The status indicator shows whether the local OpenTelemetry Collector is running.
• Iniciar, Parar, and Reiniciar control the managed collector process.
• The live log keeps the latest 300 lines and follows new output while you are near the bottom.
• Atualizar, Ir ao fim, and Limpar help inspect startup, health checks, and export errors.

The Evals tab centralizes the @dynatrace-oss/dt-evals workflow. Prompt capture must be enabled in Configurações before Evals can be activated; both toggles stay synchronized.

• Instalar dt-evals installs the CLI.
• Abrir wizard de configuração connects the tenant and configures the LLM judge.
• Rodar Evals selects evaluators, time range, sample size, and execution mode.
• Validar setup checks configuration and connectivity.

Interactive actions open in the integrated terminal so progress and prompts remain visible.

Open Notebooks in your tenant and run these queries.

Claude Code — prompt stream

fetch spans, from:now()-24h
| filter service.name == "claude-code"
| filter startsWith(span.name, "chat")
| filter isNotNull(`gen_ai.prompt`)
| fields
    timestamp,
    Developer     = `user.email`,
    Model         = `gen_ai.request.model`,
    Duration      = duration,
    Input_Tokens  = `gen_ai.usage.input_tokens`,
    Output_Tokens = `gen_ai.usage.output_tokens`,
    Prompt        = `gen_ai.prompt`,
    Response      = `gen_ai.completion`
| sort timestamp desc

Claude Code — usage by developer

fetch spans, from:now()-24h
| filter service.name == "claude-code"
| filter startsWith(span.name, "chat")
| summarize
    Requests      = count(),
    Input_Tokens  = sum(toLong(`gen_ai.usage.input_tokens`)),
    Output_Tokens = sum(toLong(`gen_ai.usage.output_tokens`)),
    Total_Tokens  = sum(toLong(`gen_ai.usage.input_tokens`) + toLong(`gen_ai.usage.output_tokens`)),
    Avg_Duration  = avg(duration)
  by: Developer = `user.email`
| sort Total_Tokens desc

Claude Code — usage by model

fetch spans, from:now()-24h
| filter service.name == "claude-code"
| filter startsWith(span.name, "chat")
| summarize
    Requests     = count(),
    Total_Tokens = sum(toLong(`gen_ai.usage.input_tokens`) + toLong(`gen_ai.usage.output_tokens`)),
    Avg_Duration = avg(duration)
  by: Model = `gen_ai.request.model`
| sort Total_Tokens desc

Claude Code — volume over time (7 days)

fetch spans, from:now()-7d
| filter service.name == "claude-code"
| filter startsWith(span.name, "chat")
| summarize
    Requests = count(),
    Tokens   = sum(toLong(`gen_ai.usage.input_tokens`) + toLong(`gen_ai.usage.output_tokens`))
  by: bin(timestamp, 1h)
| sort timestamp asc

Claude Code — cost estimate per developer

fetch spans, from:now()-30d
| filter service.name == "claude-code"
| filter startsWith(span.name, "chat")
| summarize
    Total_Cost_USD = sum(toDouble(`gen_ai.usage.cost_usd`)),
    Requests       = count()
  by: Developer = `user.email`
| sort Total_Cost_USD desc

Claude Code — tool calls

fetch spans, from:now()-1h
| filter service.name == "claude-code"
| filter startsWith(span.name, "claude.tool.")
| fields timestamp, span.name, claude.tool_input, claude.tool_response, claude.duration_ms
| sort timestamp desc

GitHub Copilot Chat spans

fetch spans, from:now()-1h
| filter service.name == "copilot-chat"
| filter isNotNull(`gen_ai.conversation.id`)
| fields timestamp, span.name, gen_ai.request.model,
         gen_ai.usage.input_tokens, gen_ai.usage.output_tokens
| sort timestamp desc

The isNotNull(gen_ai.conversation.id) filter excludes Copilot's internal background calls (e.g. orchestration using gpt-4o-mini) that have no conversation context and are covered by the flat Copilot subscription — not billed per token to you.

Workaround for users with their own API key: configure Cursor to use a custom endpoint in Cursor Settings → Models → Custom pointing to a local OpenAI-compatible proxy that emits OTel spans. This is not included in this extension today.

This is a known gap — native Cursor AI observability requires Cursor to add OTel support to their platform.

vscode-dt-ai-observability/
├── src/
│   ├── extension.ts      — entry point, activation and configure flow
│   ├── collector.ts      — OTel Collector process lifecycle, port auto-discovery
│   ├── downloader.ts     — platform detection, binary download and cache
│   ├── claudeHooks.ts    — Claude Code hook script (Python) + auto-install
│   ├── configPanel.ts    — three-tab configuration and operations webview
│   ├── evals.ts          — dt-evals installation and terminal workflows
│   └── statusBar.ts      — VS Code status bar indicator
├── resources/
│   └── otel-collector.yaml  — Collector config (uses env vars for credentials)
├── scripts/
│   └── package.sh        — builds the VSIX
├── README.md
└── package.json