OpenTag just dropped into the CopilotKit ecosystem, and it's exactly what teams have been asking for: a self-hosted AI agent that lives right inside your Slack channels. Instead of paying per-seat fees or dealing with vendor lock-in, you run the whole stack yourself — bringing your own model (Anthropic's Claude or OpenAI's GPT) and wiring it to whatever tools your workflow needs. The premise is simple. Drop @OpenTag into any thread, ask it to summarize a conversation, generate a report, or file a ticket in Linear, and it'll do the work inline. The agent renders rich results — tables, breakdowns, even bar charts — directly in the Slack thread using CopilotKit's generative UI components. Before taking destructive actions like creating issues or posting updates, OpenTag waits at an "Approve" gate for human confirmation. It's Claude for your workspace, except you own the runtime and there's no subscription model breathing down your neck. Under the hood, you're running two separate processes: the agent backend on port 8200 (the LLM plus optional MCP tools) and the bot that maintains the Slack connection. Setup requires three secrets — SLACK_BOT_TOKEN, SLACK_APP_TOKEN, and either OPENAI_API_KEY or ANTHROPIC_API_KEY depending on which model you're proxying to. The whole thing starts with pnpm install followed by two filtered commands from the CopilotKit monorepo root. A detailed walkthrough lives in setup.md for anyone who wants to see every step before committing to the install.

Customization Is the Whole Point

What makes OpenTag worth your time over just using Claude directly is how hackable it is. The agent's behavior lives in a single system prompt file (runtime.ts) — rewrite that and you've got a completely different assistant without touching any infrastructure code. Copy the app/ directory to bootstrap your own bot on other platforms since createBot accepts an array of adapters for multi-platform support. Currently Slack works out of the box, with Discord, Telegram, and WhatsApp integrations available through the same codebase.

The Managed Service Looms

Let's be real: running chat agents in production is not trivial. CopilotKit acknowledges this upfront on their GitHub page — "We won't lie to you, though. Setting up hosting for chat agents is not easy." That's precisely why they're building a managed version on their Intelligence platform with durable threads, persistence, hosted inspection tooling, and what they're calling Continuous Learning from Human Feedback (think RLHF but for your team's specific workflows). The waitlist is live now if you'd rather pay someone else to handle the ops overhead. OpenTag represents a shift toward owning your AI infrastructure rather than renting it. For shops with strong DevOps maturity and strict data governance requirements, this kind of self-hosted approach eliminates compliance headaches that come with sending sensitive Slack conversations to third-party APIs. The trade-off is operational complexity — you're on the hook for scaling, monitoring, and keeping the runtime healthy.

Key Takeaways

  • OpenTag runs in Slack threads as a @mentionable bot, rendering tables, charts, and summaries inline with generative UI
  • Self-hosting requires two processes (agent backend + bot) plus three environment secrets
  • System prompt lives in one file — swap it out to change agent behavior without touching infrastructure
  • Discord, Telegram, and WhatsApp adapters exist in the same codebase using CopilotKit's multi-platform SDK

The Bottom Line

OpenTag is a solid proof-of-concept for teams that want Claude-level AI assistance inside Slack without surrendering control or paying per-seat premiums. The catch? You better have someone comfortable managing LLM runtimes, because the managed service isn't here yet and self-hosting chat agents at scale is genuinely hard. Worth watching — just don't expect this to be a plug-and-play solution out of the box.