Playground

Try actions in natural language before wiring code.

The AI Playground runs any action from your linked accounts using natural language — no code, no setup. Use it to confirm an account works and see which actions are available before building.

1. Select accounts

   Pick one or more linked accounts from the dropdown.

2. Configure actions (optional)

   Expand Actions to enable or disable specific actions. Selection locks once the conversation starts.

3. Chat

   Try a prompt like the ones below — each tool call and its response appear in the chat log.

   What tools can I use?
   

   List the first 10 employees
   

Start with “What tools can I use?” to see the available actions for your selected accounts.

Model selection

Click the model selector at the bottom right to switch models:

Model	Recommendation
Claude Haiku 4.5	Recommended. Best balance of speed and tool calling accuracy. Selected by default.
Claude Sonnet 4.6	Use for complex multi-step reasoning tasks.
GPT-5.5	OpenAI alternative for comparison testing.
GPT-5.4 Mini	Lighter, faster OpenAI model.
GPT-5.4 Nano	Smallest OpenAI model for low-cost runs.

StackOne Tool Calling Model (Beta)

StackOne is fine-tuning a specialized model optimized for SaaS tool calling, at lower cost than leading models. Contact [e-mail] to join the beta.

Defender

The configuration sidebar has a StackOne Defender toggle that scans tool call responses for prompt injection before the model sees them. It’s a per-request override: it applies to the Playground session only, whatever your project-level setting is, and defaults to on. Use it to see how your agent behaves with Defender on or off without changing your project configuration.

With Defender enabled, every tool call in the chat log carries a risk icon next to its status badge:

Icon	Meaning
	Low risk — response passed Defender’s checks
	Medium risk — flagged but not blocked
	High or critical risk — blocked from reaching the model when Block High Risk Content is on in Defender
no icon	Defender did not scan this response (toggle off, size limit exceeded, or no Defender metadata)

Hover any icon for the precise risk level. When a call is blocked, it shows a Blocked status in the chat log and the model receives an error in place of the tool response, which the agent can retry, skip, or surface like any other tool error.

Defender

Detection modes, thresholds, large-response behavior, and SDK overrides.

Common errors

No accounts visible

Check the account is linked on the Accounts page, and that actions are enabled on its connector profile.

Slow responses

Switch to Claude Haiku (the fastest option) or reduce the number of selected accounts.

Tool not found

Verify the action is enabled on the connector profile, then check Request Logs for the failing call.

Connect your own client

The Use Setup button shows the MCP config for your accounts (endpoint, authentication header, x-account-id) so you can move from testing to your own agent. See MCP.

The Playground is a testing tool, not a production agent. Build your agent with MCP or the Agent SDK.