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

# Configuration

> The API key, every setting, and exactly what init() does at startup.

## The API key

Everything starts with an Eldros API key (format `agt_...`). Generate it from the Eldros platform dashboard: open your agent, go to **Configuration**, and under **Trace API Keys** give the key an optional label (e.g. `production`) and click **Generate new key**.

<Frame caption="Your agent's Configuration page → Trace API Keys → Generate new key">
  <img src="https://mintcdn.com/eldros/7xelfYg-q2j7-sc9/assets/eldros-api-key.png?fit=max&auto=format&n=7xelfYg-q2j7-sc9&q=85&s=d58545fdbde3a639f9f293dca44af0d9" alt="Trace API Keys section on the agent's Configuration page in the Eldros dashboard" width="2446" height="778" data-path="assets/eldros-api-key.png" />
</Frame>

<Warning>
  A new key is **shown once and cannot be retrieved later** — only a masked prefix is stored. Copy it into your secret manager immediately. And treat it like any credential: environment variable or secret store, never committed to source control.
</Warning>

Set it as an environment variable:

```bash theme={null}
export ELDROS_API_KEY=agt_...
```

`init()` uses the key to resolve your trace configuration from the Eldros platform at startup — you never configure endpoints, headers, or backends yourself, and rotating infrastructure on our side never requires a client change.

### One key per agent

Generate a **separate key for every agent you onboard** — the key *is* the agent's identity. That's what keeps each agent's traces and conversations isolated from one another, and it's how Eldros knows which agent a simulation run or a judged conversation belongs to. Sharing one key across agents merges their telemetry into a single stream that can't be cleanly separated afterwards.

## All settings

Explicit `init()` kwargs always win over environment variables; environment variables win over defaults.

| `init()` kwarg                | Environment variable                 | Default              | What it does                                                                                        |
| ----------------------------- | ------------------------------------ | -------------------- | --------------------------------------------------------------------------------------------------- |
| `api_key`                     | `ELDROS_API_KEY`                     | —                    | your Eldros key (`agt_...`)                                                                         |
| `api_base_url`                | `ELDROS_API_BASE_URL`                | Eldros cloud         | only set if directed (dedicated deployments)                                                        |
| `traffic_type`                | `ELDROS_TRAFFIC_TYPE`                | `"prod"`             | `"prod"` for live traffic, `"eval"` for internal testing — keeps test noise out of production views |
| `instrument_openai`           | `ELDROS_INSTRUMENT_OPENAI`           | extra installed → on | force LLM auto-instrumentation on/off ([details](/docs/tracing))                                         |
| `instrument_anthropic`        | `ELDROS_INSTRUMENT_ANTHROPIC`        | extra installed → on | same, for Anthropic                                                                                 |
| `instrument_claude_agent_sdk` | `ELDROS_INSTRUMENT_CLAUDE_AGENT_SDK` | extra installed → on | same, for the Claude Agent SDK                                                                      |
| `capture_content`             | `ELDROS_CAPTURE_CONTENT`             | `true`               | record prompt/response text on LLM spans ([privacy](/docs/tracing#content--privacy))                     |
| `debug`                       | `ELDROS_TRACE_DEBUG`                 | `false`              | also print spans to the console                                                                     |
| `enabled`                     | `ELDROS_TRACE_ENABLED`               | `true`               | master switch — `false` makes the whole SDK a no-op                                                 |

Boolean environment variables accept `1/true/yes/on` (anything else is false).

## What init() does at startup

```python theme={null}
eldros_sdk.init(traffic_type="prod")
```

1. Resolves settings (kwargs → env vars → defaults).
2. Makes **one brief blocking call** to the Eldros platform to fetch your trace configuration — call `init()` at import time or in your startup hook, **before your event loop starts**, not inside it.
3. Wires up OpenTelemetry and auto-instruments the LLM libraries whose [extras are installed](/docs/tracing).

Call it once per process; repeat calls are ignored with a warning.

## Failure behavior — designed to never take you down

* **No API key / resolution fails:** the SDK logs a warning and degrades — spans are still created but not exported. Your agent runs exactly as before.
* **`enabled=False`** (or `ELDROS_TRACE_ENABLED=false`): everything becomes a hard no-op — the documented kill switch for incidents or environments where telemetry must be off.
* **Not initialized at all:** every SDK call (`turn()`, instrumentation) is a safe no-op. It is safe to leave Eldros code in paths that sometimes run without `init()`.