CopilotKit Open Sources Channels SDK: An MIT Licensed Library That Runs Any AG-UI Agent Inside Slack And Microsoft Teams
MarkTechPost Asif Razzaq
CopilotKit open-sourced Channels SDK, letting any existing AI agent drop straight into Slack or Teams. No rewrite needed — you just plug in your agent, and it renders native buttons and cards.
CopilotKit shipped something this week that solves a problem most agent builders quietly dread: getting a working AI agent out of a demo notebook and into a place where actual humans already hang out, like Slack. The new Channels SDK, released under MIT license as version 0.5.0, treats messaging apps as a transport layer rather than a new brain to build. If your agent already speaks AG-UI — CopilotKit's own agent-user interaction protocol — it works with LangGraph, CrewAI, Mastra, Pydantic AI, Google ADK, or a raw HTTP endpoint you hacked together yourself, without touching your model or your tool logic.
The installation is almost insultingly simple: two npm packages, a few lines of config, and a channel object pointed at Slack or Microsoft Teams, both listed as early access. Discord, Telegram and WhatsApp adapters are already sitting in the package for later. But the interesting engineering choice isn't the install, it's the rendering pipeline. Messages get authored in JSX, get flattened into a serializable format, and then get translated into Block Kit for Slack or Adaptive Cards for Teams. That's a real UI, with buttons and interactive handlers wired through an action store, not a wall of text stuffed into a webhook.
CopilotKit split the responsibilities cleanly. Your side owns the agent, the model credentials, the tools, the business logic, and a long-running Node process — serverless won't cut it here since the connection has to stay alive. Their side, called CopilotKit Intelligence, handles the Slack and Teams platform credentials, ingress, reconnects, and health checks, and every setup requires an Intelligence API key whether you go the managed or direct route. The runtime contract is deliberately opinionated too: there's no channel.start() method. The listener owns activation, and you're expected to await a channels.ready() call so a broken config fails loudly at startup instead of silently rotting in production.
What makes this more than a wrapper is the feature set aimed squarely at real workflows instead of chat gimmicks. Agents can render generative UI mid-conversation, pause for human approval before doing something risky, call tools to file a Linear issue or open a pull request, and carry memory across platforms through a persistence layer. OpenTag's implementation shows the approval pattern in practice — every write to Linear or Notion gets intercepted before the MCP request fires, and a confirm_write has to clear before anything actually mutates. Reads keep flowing without interruption, which is the kind of detail that separates a demo from something a team would trust in a live channel.
The requirements are unglamorous but sensible: Node.js 22 or newer, an ESM project, and a process that stays up. Nothing exotic, just the plumbing that production software actually needs.
My take
This is the boring, useful kind of open source release that doesn't get much hype but quietly removes a real barrier — most agent teams have been hand-rolling Slack integrations for months, badly. Betting on AG-UI as the connective tissue is smart because it avoids locking anyone into CopilotKit's own agent framework, which is more than can be said for a lot of vendor SDKs dressed up as open standards. The catch is that every path still routes through CopilotKit's Intelligence API key, so calling this fully open is generous — it's open at the edges and metered in the middle.
Read more about this at: MarkTechPost
