What changed
Vercel now supports running Anthropic’s Claude Managed Agents through its Chat SDK, a combination that lets you deploy a working chatbot, like a research assistant that lives in a Slack channel, without writing your own server to manage the conversation.
The two pieces come from different companies. Claude Managed Agents is Anthropic’s own hosted service: instead of you writing the code that loops the AI model through steps, calls tools, and remembers what happened earlier in a conversation, Anthropic runs that loop for you on its own servers. Chat SDK is Vercel’s toolkit for wiring an AI agent into a chat interface, including the plumbing needed to make it work inside Slack, Microsoft Teams, Google Chat, Discord, WhatsApp, and other messaging apps, described as “30+ platforms,” without rewriting the whole thing for each one.
Put together, Vercel says you get a chatbot with one persistent conversation per Slack thread, replies that stream in as the model writes them (rather than appearing all at once when it’s done), and a live activity feed that shows which tools the agent is calling and what it’s searching for while it works. Because Claude Managed Agents already stores the conversation on Anthropic’s side, Vercel says you don’t need to run your own database to keep track of chat history. Vercel’s own path to try it is a one-click template: deploy a Slack research agent from a prebuilt project.
Why it matters
Vercel says “Anthropic stores the conversation for you”. Before routing sensitive conversations through the bot, check Anthropic’s retention and compliance limits.
Claude Managed Agents is currently in public beta, and Anthropic’s own documentation is explicit that it’s “stateful by design”: sessions are long-running, retain state across pauses, and store your conversation history, sandbox activity, and outputs on Anthropic’s servers so features like Vercel’s “no database to run” setup work at all. Anthropic says Managed Agents is not eligible for Zero Data Retention, which avoids retaining request data, or HIPAA Business Associate Agreement coverage. That agreement sets obligations for a vendor handling protected health information.
Persistent history makes the setup convenient, but also places it outside that retention opt-out. Public documentation may fit those terms, though account access and tool permissions still need review. If you’re picturing it summarizing internal customer conversations, financial figures, or anything under a compliance obligation, that’s worth ruling in or out before you connect it to a real Slack workspace, not after.
You’ll also need your own Claude API key to run this at all. Access is enabled by default on Anthropic API accounts, but Claude API usage is billed by the token (the small chunks of text the model reads and writes, which is how most AI companies measure usage), not something bundled free into a Vercel plan. Check measured session usage in your Anthropic billing dashboard before estimating ongoing costs.
Who should care
The template still requires code configuration. A handler, the code that responds to chat events, connects the agent to Slack or another app. It’s most useful if you want an AI assistant living inside a tool your team already uses daily, like Slack, instead of a separate chat window nobody remembers to open. It matters less if you’re happy with a standalone chatbot experience, or if the conversations you’d route through it involve information you can’t put in front of a third-party AI vendor’s servers without a data-handling review first.
What builders should do next
Before deploying anything real, check Anthropic’s data retention terms for your account. Decide which topics the bot may handle under its “stateful by design” storage model.
Then test the template with one non-sensitive request you can verify, such as asking it to summarize a public documentation page. Check whether each factual claim traces to a source in the activity feed and whether the reply arrives at a reasonable pace for the task. If it invents sources or the activity feed leaves the work hard to inspect, investigate the setup before connecting real team conversations.
End of article