---
title: "Use the Autohand Playground Docs"
source: https://docs.autohand.ai/guides/playground
---

# Use the Autohand Playground

Compare one software-delivery task across up to three model windows, inspect each result beside its prompt, and carry the active task into your application with current SDK code.

[Open Playground](https://console.autohand.ai/playground) [SDK quickstart](https://docs.autohand.ai/agent-sdk/quickstart)

## Before you run

-   Sign in at [console.autohand.ai](https://console.autohand.ai/) and choose the intended account.
-   Use evidence you are allowed to send to the selected model. Do not paste raw secrets, access tokens, or customer data you do not need.
-   Start with Fantail for a fast, focused loop. Choose Moa when the task needs broader context or deeper reasoning.
-   The browser uses your Console session. Copied SDK examples use `AUTOHAND_API_KEY` in the environment, never a raw key embedded in source.

 [Download the compare-models walkthrough](https://docs.autohand.ai/media/guides/playground-compare-models.webm).

In 7 seconds: choose a bug-fix example, add a third window, and run the same evidence across all three. The recording uses a local fixture, so it sends no model request.

1.  **Examples menu**

    Choose Goal, Deep research, Bug fix, Review, Test, Migration, Release, or Security. The selected task fills every window.

2.  **Comparison windows**

    Start with Fantail and Moa. Add or duplicate a window to compare up to three model and parameter combinations.

3.  **Prompt and output**

    Edit each system and user message independently. Its result remains in the same vertical lane.

4.  **Run, History, and SDK code**

    Run one window or all windows, reopen saved runs, and copy syntax-highlighted code without leaving the workbench.

## Compare a task across model windows

1.  **Choose an example.** Open the examples menu. The groups cover planning and research, building and debugging, and review and release. Selecting an example applies the same complete task to every window.
2.  **Set up the comparison.** Playground starts with Fantail and Moa. Select **Add window** to add a third variation, or open a window's actions menu and select **Duplicate window**.
3.  **Choose each model.** Use Fantail for a fast, focused loop. Use Moa for broader context or deeper reasoning. Each window keeps its own model and parameters.
4.  **Edit the messages.** In System, define the role, evidence rules, and expected answer. In User, include the exact diff, error, contract, rollout facts, or test output.
5.  **Adjust parameters when needed.** Select the sliders button to choose streaming, temperature, output limit, sampling, response format, and cache behavior. Moa exposes reasoning effort; Fantail omits it because the model does not accept that parameter.
6.  **Run and compare.** Select **Run** for one window, or **Run all** (Cmd/Ctrl Enter) for every window. Compare the outputs against the supplied evidence instead of treating the first response as approval.

 [Download the request-controls walkthrough](https://docs.autohand.ai/media/guides/playground-request-controls.webm).

In 5 seconds: turn streaming off, set reproducibility and JSON-output controls, then confirm that the syntax-highlighted request is the exact body Run will send.

### A reusable review prompt

```text
Review this change for correctness, security, and rollout risk.

Evidence:
- Intended behavior: [state the contract]
- Changed code: [paste the smallest relevant diff]
- Existing tests: [name or paste results]
- Deployment constraint: [state the real boundary]

Return:
1. Confirmed findings, ordered by severity
2. Assumptions that still need evidence
3. The smallest tests required before approval
```

## Reopen private runs for 30 days

After a successful run, Playground saves the exact validated request JSON, response, and run metadata privately to the current member inside the current account.

-   Each saved run receives a fixed expiry exactly 30 days after it was created.
-   Expired runs stop appearing automatically and are deleted by scheduled cleanup.
-   Opening a run does not extend its expiry.
-   Other members in the same account cannot open your Playground history.
-   Select **Delete** beside a run and confirm to remove it sooner.

To restore a run, first select the destination window. Select **History**, then select the run title. Playground restores the original model controls, prompts, result, and run evidence into the active window. Editing the restored prompt does not change the saved record; running it again creates a new record with its own expiry.

 [Download the history walkthrough](https://docs.autohand.ai/media/guides/playground-history-restore.webm).

In 6 seconds: run the comparison, open History, and restore a saved result into window A. Opening a run does not extend its expiry.

## Move the task into your application

Select **SDK code** to generate the active window's prompt for REST and each documented SDK. The REST tab reproduces the visible request JSON exactly; SDK tabs use each SDK's real lifecycle and identify controls that are available only on the completion API. Syntax highlighting and line numbers make every example easier to review before copying.

 [Download the SDK-code walkthrough](https://docs.autohand.ai/media/guides/playground-sdk-code.webm).

In 6 seconds: open SDK code, switch from TypeScript to Python and Go, then copy the highlighted example. Each tab uses that SDK's native lifecycle.

1.  Create or choose a dedicated key under [Console API Keys](https://console.autohand.ai/api-keys).
2.  Store the raw value as `AUTOHAND_API_KEY` in the server or local environment that launches your app.
3.  Select REST, TypeScript, Python, Go, Java, Swift, Rust, Ruby, C#/.NET, or C++.
4.  Run the displayed install command, copy the code, and keep the model input appropriate for your application.
5.  Remove any temporary key after the integration check if the application will not continue using it.

## Troubleshooting and checkpoint

| Symptom | What to check |
|---|---|
| History does not load | Refresh Console, confirm you are still signed in, and verify the selected account. A run remains usable even if its background history save fails. |
| A run is missing | Only successful runs are stored. The run may also have expired after 30 days or been deleted. |
| Windows are answering different tasks | Choose the example again to apply one prompt to every window, then make only the model or parameter change you intend to compare. |
| The answer assumes missing facts | Add the smallest relevant diff, logs, contract, and verification result. Tell the model to separate confirmed findings from assumptions. |
| Copied code cannot authenticate | Keep AUTOHAND_API_KEY server-side, confirm the key is active, and follow the API-key setup guide. |

You are finished when the result answers the supplied evidence, the run appears in History with its expiry, restoring it recovers the original state, and your chosen integration uses the documented SDK lifecycle.