The fastest way to investigate Atmos Pro today is to point Claude Code at the workspace's MCP server and let it chase drift, summarize failures, and surface remediation plans. That works, but it means leaving the product to do it. Atmos AI brings that same experience into the dashboard.

Click Ask AI in the header, or select stacks on the Stacks page and choose Ask AI ( stacks). A side panel opens with a workspace-scoped agent that can call the same investigative tools the MCP server exposes — workflow runs, logs, AI summaries, instances, drift, deployments, pull requests, audit log, and documentation search.

The default model is Claude Sonnet 4.6 with extended thinking enabled. As the agent works, the panel shows three classes of streaming content:

Each message has a copy icon, and a Copy conversation button in the panel header serializes the entire thread — including the tool-call breadcrumbs — as Markdown you can paste straight into Claude Code or any other agentic editor.

Opening the chat from the Stacks multi-select pre-fills the agent's context with the selected instance IDs so its first tool call is already scoped. Every session is workspace-scoped and persisted: the session ID lives in the URL, so a hard refresh restores the thread, and /chat lists prior conversations.

Atmos AI reuses the existing Atmos Pro MCP server in-process. There is no second tool surface to maintain and no extra OAuth dance — the chat agent calls the same code that serves /mcp, with the workspace user's role permissions enforced at the data layer. The MCP server remains available for users who prefer to drive it from Claude Code.

Atmos AI is included in every paid plan and can be disabled per-workspace.