Playground

Playground is a private browser conversation with agents running on your daemons. Open it from an agent page for a single-agent session, or use the Home composer to bring several agents into the same conversation.

What you getCopy Anchor Link

Playground creates a real session on the agent's daemon, using the same runtime, workspace, and tools as a channel conversation. Its output streams into the same kind of session transcript.

You can interrupt a turn when needed. When the agent's What people can do in chat is Talk, approve and change settings, you can also change supported runtime settings for the next turn without altering the agent's saved configuration. A tool call that needs approval waits among the pending requests above the composer. The agent's editors can answer it at any level, and from Talk and approve tool calls up, so can you as the conversation's owner, even if you can only view the agent.

When you ask the agent to read or post in a chat conversation, it does so for you, by the same rule as in chat; see Reading and posting in other conversations. In a workspace where you linked your chat account under Your profile → Sign-in methods, such as your Slack account, it judges you by that account: a public channel needs you to be a full member, and a private channel or group DM needs you to be a member. Anywhere else, including Telegram and Discord, whose accounts can't be linked, you reach only conversations whose sessions you can open under Sessions.

Your own conversation with an agentCopy Anchor Link

Where the agent's daemon runs Auto, the agent page's Playground button, and a message to one agent from the Home composer, open your own conversation with that agent. Only you can open it, and it is the same conversation every time. A coordinator answers each message there: it replies to a quick question itself and opens a task for anything that needs work. Each task continues on its own session page, which the conversation links to, and you can keep writing to the task there. New session in the Sessions panel opens an ordinary Playground conversation instead; use it for the setup work below.

Configure AgentConnect from a conversationCopy Anchor Link

In a Playground conversation with a single agent, that agent can read your setup and make changes for you with the same tools as the MCP connector. It acts as you, so your role and visibility apply. Each change appears as an approval card above the composer with its exact arguments, and nothing is applied until you choose Approve and run. The agent is then told the result and continues.

Start with the built-in AgentConnect agent for setup work; it comes with skills for the platform and for creating agents. Any agent can also change its own configuration and workspace this way, but it cannot delete itself.

These tools are not available in a conversation with several agents. Adding an agent to a conversation removes them for the rest of that conversation, so start a new single-agent conversation for setup work.

The agent never asks for credentials in chat. To create an agent, connect an integration or code host, or add a secret, MCP server, or skill, it opens the matching console dialog, prefilled where it can, for you to finish. Trying a Decision on a sample message is not available here; use Try with an example in the Decision editor.

Attach filesCopy Anchor Link

Use the composer's attach button to add files to a message. An image goes to the agent with the message itself. Any other file uploads to your deployment's object store first and appears as a chip above the composer; the message sends once every upload has finished. The agent receives the file's name and a short-lived download link, and fetches the file into its workspace when it needs it.

• A message can carry up to 4 files, each within your deployment's upload limit; the composer tells you when a file is over it.
• If your deployment has no object store configured, only images can be attached.
• A message with files is never folded into a turn that is already running. It waits and runs as its own turn once the current one ends.

Work with several agentsCopy Anchor Link

Add agents before the first message from the Home composer, or add them to an existing Playground conversation.

• A message without an @AgentName goes to every participant. Each agent may answer or stay silent.
• An explicit @AgentName narrows that turn to the named participant or participants.
• Files go only to the agents a message addresses. The other participants see that files were attached, by name, but cannot open them.
• Every agent keeps its own configuration while the browser combines their activity into one conversation.

When the session is savedCopy Anchor Link

Opening Playground starts with an empty, temporary view. Your first message saves the private conversation under Sessions, where you can reopen it later.

TipsCopy Anchor Link

• Test permission modes here before using an agent in a shared channel — try Plan or Ask for approval first, watch what it wants to do, then relax. Switching the mode from the composer needs Talk, approve and change settings; otherwise change it in the agent's Edit. Before the agent joins a public group, set it to Talk to the agent, so nobody there can approve its tool calls.
• The Playground respects the agent's workspace. For a new single-agent conversation backed by GitHub, choose whether its first turn uses an isolated worktree without changing the agent's saved default.
• If the composer reports an agent as unreachable, check its daemon is online — Playground sessions run on the participating daemons, not in the cloud.