Skip to main content
Playground can use the tools your organization has enabled and your account has connected. It uses your signed-in session for tools and the selected project API key for each model round. A separate MCP token is only needed for external agents.

​
Start a tool conversation

  1. Enable tools and connect your accounts.
  2. Open Playground, choose a model with function-tool support, and select the project key.
  3. Choose Select tools beside the message input and select connected accounts.
  4. Ask for a task that uses one of those connections.
  5. Review each proposed action’s name and arguments, then approve or decline it.
Changing the selected tools starts a new conversation. A disconnected personal account links to Agent tools; an administrator manages missing shared connections.

​
Search and approve actions

Playground gives the model aixy_search_tools and aixy_execute_tool. Metadata searches run automatically and return compact JSON from the search service. The model reads complete parameters for its chosen action through the same search helper; small results can include parameters immediately. Searching and reading parameters do not execute the action. The model sees semantic names such as github.get_me. Definition revisions and approval metadata remain outside model history. Small supporting actions can include parameters in the same response. Every external action, including a read, needs your approval. Approval is tied to the discovered connection revision. If configuration changes, start a new conversation and discover the action again. Declining sends a declined result to the model without performing the external action. Review results and duration in the action details. Multiple proposals run sequentially after approval. Tool results can become input to subsequent model rounds, which use the same project controls and may incur additional provider cost.

​
Stop and recover

Stop cancels the turn and pending approvals. An action already sent may still complete. An interrupted execution can show Outcome unknown: inspect MCP Activity and the external service before deciding whether to repeat it. Playground never retries external actions automatically. An initial model HTTP error preserves the draft. An interrupted stream or failed continuation after an approved action requires New conversation so completed actions are not replayed. A turn allows up to sixteen model rounds and twenty proposed external actions, with up to 20 selected connections.

​
Move to an external agent

Use Agent setup to open the connection instructions. A tool conversation involves discovery, approval, execution, and model continuations, so configure the client using Connect an agent. Conversation text and tool results remain in the page’s memory and clear on navigation, reload, or a new conversation. Explicitly enabled Activity capture continues to apply to model traffic. MCP Activity retains execution metadata without arguments or results. Discovery retains only selected action IDs and revision metadata in an expiring conversation context, separately from the transcript; it stores no prompts, arguments, results, or credentials. Discovery returns compact JSON with semantic names. The Playground sends the same name for approved execution. Names use the namespace.action format, such as github.get_me, and refer to the action selected in the current conversation context.