Skip to main content
Start by sending one small, synthetic run to your workspace. You will see its input, tool result and output in Telemetry. Then replace the example with your agent.
This example calls no language model and performs no business action. It sends data to Sentrial, where your workspace’s processing and usage settings apply. Receiving a trace does not establish signal accuracy or prove that an agent completed its task.

1. Install the SDK

For guided Node.js setup, run this in your agent project:
The wizard checks workspace access and creates an integration example. Run that example to send a trace, then check delivery:
See the CLI guide for unattended setup. If you prefer a self-contained first trace, follow the manual steps below. Use Python 3.9+ or Node.js 18+. These native examples work with the published Python 0.7.0 and TypeScript 0.6.0 SDKs; no framework or additional model key is required.

2. Configure your workspace

Sign in, choose the workspace you want to monitor, and open Settings → API Keys. Supply its key through your terminal or secret manager as SENTRIAL_API_KEY; keep it out of source files and Git. The key needs sessions:write and events:write to send the example. Reading sessions through the CLI additionally requires sessions:read and events:read. An ingestion key does not grant investigation access. The default API is https://api.sentrial.com. Set SENTRIAL_API_URL only when using a different Sentrial environment, such as your staging API. A key belongs to one workspace and environment.

3. Send a first trace

Save the example for your runtime. It calculates 2 + 3, records the tool’s actual input and result, and flushes telemetry before exiting.
Run the file:
The script prints a session ID. The success value describes this example’s execution; it is not a model judgment about business success.

4. Inspect the evidence

Open Telemetry in the same workspace and inspect the session with the printed ID. Check that its input is What is 2 + 3?, the add tool received both arguments, the tool returned 5, and the final output is 5. If it is missing, confirm the workspace, API origin and write scopes first. An acknowledged session ID is not proof that every event arrived. Inspect delivery warnings and missing evidence before trusting the trace. Ingestion options and coverage explain the available paths. For terminal inspection, the published Node CLI supports:
Use the same API origin and a key with read scopes. See the CLI reference for command/version requirements and pagination limits.

5. Connect your agent and inspect a failure

Replace the calculation with your agent’s work. Preserve each real tool result, including errors, pending work and retries; don’t replace it with success: true. Use a supported framework integration to capture its LLM calls and nested tool steps, or record them explicitly through the native SDK. In your workspace, Signals → Define a signal describes the behavior you want to watch. Review decisions shows captured evidence, saved predictions and corrections, including missing or uncertain results. Evidence capture and evaluation must be configured for that workflow; installing an SDK alone does not evaluate every run or automatically ship a fix.