Try tools in the playground

Chat with a model that calls your MCP tools through the gateway, from the console, before you connect a client.

The playground is a chat in the console where a model can call the MCP tools your organisation has registered. Pick a model and some tools, ask something that needs them, and watch each tool call as it happens: its name, arguments, result, latency and any error. It is the quickest way to check that a server works through the gateway, and to see what an agent would see, before you set up an MCP client.

The playground uses the MCP gateway, so it is available on plans that include it. See Billing, plans and credits.

Who can use it

Using the playground needs the playground.use permission. Owners and admins have it; members do not, unless a custom role gives it to them. See Roles and permissions. Without it, the Playground tab does not appear.

Start a chat

Go to Gateway › Playground.
Under Setup, choose a Model. The list is the models the gateway can reach for your organisation, through your provider keys or credits.
If your organisation scopes requests by team, choose a Team. Calls count against its budget.
Under Servers and tools, select the servers, or individual tools, the model may use.
Ask something that needs a tool, and click Send.

The model sees only the tools you selected. If it asks for any other tool, the call is not made, and the chat says so.

You can also start from Gateway › MCP servers. List tools on a server's row shows the tools the server offers through the gateway, and which of them are visible to you, with the reason when one is hidden. From there, Open in playground opens the playground with that server's tools selected.

Approve each tool call

Ask before each tool call is on when you open the playground. Tools can change real systems, so with this on nothing runs until you approve it. Each tool call the model asks for appears as a card, Awaiting approval:

Approve runs it. When the model asks for several calls at once, they run once every pending call is decided.
Skip does not run it. The model is told it was not run.

Turn the switch off to let tool calls run as soon as the model asks for them. Stop ends the current message; a tool call that was already running may still have run.

Limits

Each message you send can lead to several rounds of tool calls. The playground stops a message after:

8 tool rounds,
16 tool calls, or
120 seconds.

A chat can select up to 64 tools, and each member can send up to 10 messages a minute. When a limit stops a message, the chat says which one. These limits stop a runaway loop; they are not a spending control. Your budgets are.

Costs and budgets

The playground runs as you. Every model call goes through the gateway like any other request, and is paid for the way your organisation pays for requests.

Budgets apply. Your member budget, your team's budget and the organisation's budget all count the playground's model calls, and a budget that is spent stops the chat with Budget reached. Customer budgets do not apply: the playground sends no customer.
Each model call shows its tokens and cost, with View request to open it on Requests. Tool calls are listed on Requests › MCP and, like every tool call, are not priced.
Access rules apply. A tool call that an access rule refuses shows Refused by policy.
Every request and tool call from the playground carries the property playground, so you can tell them apart from your applications' traffic.

Chats are not saved. Leaving or reloading the page clears the chat. The model calls and tool calls are logged like any others, with their bodies only if your organisation keeps bodies.

Not available yet

Servers where each member signs in cannot be used in the playground yet. They are listed, marked Not available in the playground yet. Use them from an MCP client instead; see Servers that need their own sign-in.