Claude Managed Agents research analyst
Claude Managed Agents and Chat SDK power this Slack research bot. Mention it in a channel or send it a DM, and it will search and fetch sources in an Anthropic-managed sandbox before streaming a brief back into the thread.
Each Slack thread maps to one persistent Managed Agents session, so follow-ups retain the thread's prior research context. Vercel Connect owns the Slack app and credentials, while Redis preserves subscriptions, session mappings, and webhook deduplication across deployments.
Stack
| Layer | Choice |
|---|---|
| Surface | Slack via Chat SDK and Vercel Connect |
| Agent | Claude Managed Agents with Claude Sonnet 5 |
| State | Upstash Redis |
| Runtime | Next.js 16 on Vercel |
Quickstart
Requires Node.js 24+, an Anthropic API key with Managed Agents access, Vercel project, and Redis.
1. Clone and install
2. Connect Slack and Redis
Add Slack with Vercel Connect:
Add Upstash Redis from the Vercel Marketplace:
3. Create the Managed Agent
Add ANTHROPIC_API_KEY to .env.local, then run:
This one-time command creates the agent and sandbox environment, writes both IDs to .env.local, and adds them to the linked Vercel project.
4. Deploy
Invite the bot to the Slack channels where it should answer or message the bot directly.
In a thread with one human participant, follow-ups do not need another @mention. Once multiple humans participate, the bot responds only when explicitly mentioned.
Environment variables
| Variable | Required | Default | What it does |
|---|---|---|---|
ANTHROPIC_API_KEY | Yes | None | Authenticates Claude Managed Agents |
CLAUDE_AGENT_ID | Yes | None | Persistent analyst created by npm run cma:setup |
CLAUDE_ENVIRONMENT_ID | Yes | None | Anthropic-managed sandbox created by setup |
REDIS_URL | Yes | None | Stores subscriptions, deduplication, and thread/session mappings |
SLACK_CONNECTOR | Yes | None | Vercel Connect Slack connector UID |
BOT_USERNAME | No | claude-research-bot | Chat SDK bot name |
CLAUDE_DEBUG_MODE | No | false | Posts per-turn diagnostics and a Claude Console link |
Customize the analyst
The model and system prompt live in scripts/cma/lib/agent.ts.
Publish changes as a new agent version:
Anthropic pins each session to the agent version that created it, so test changes in a new Slack thread. Use npm run cma:update for prompt changes, not npm run cma:setup.
The setup disables Bash because the analyst reads untrusted web pages. If Bash were auto-approved with open network access, a malicious page could trick it into sending conversation data to an attacker.
Debug mode
Set CLAUDE_DEBUG_MODE=true to post a compact diagnostics table after every completed turn. It includes duration, model requests, token and prompt-cache usage, web searches and fetches, and a Claude Console session link.
Key files
| File | Purpose |
|---|---|
src/lib/bot.ts | Chat SDK configuration and Slack handler registration |
src/lib/research-handler.ts | Participant policy and thread/session orchestration |
src/lib/managed-agents.ts | Session validation, turn queue, event stream, and token previews |
src/lib/config.ts | Typed runtime configuration and required env validation |
src/lib/debug-card.tsx | Debug Mode diagnostics card |
src/app/api/webhooks/[platform]/route.ts | Fast webhook acknowledgement and background dispatch |
scripts/cma/cli.ts | Dispatches setup, update, and archive commands |
scripts/cma/lib/agent.ts | Model, system prompt, and tool posture |
scripts/cma/setup.ts | One-time agent and environment provisioning |
scripts/cma/update.ts | Publishes a new agent version |
scripts/cma/archive.ts | Archives Managed Agents resources |
scripts/cma/lib/env.ts | Shared local environment loading and persistence |
Cleanup
Archive the persistent Anthropic resources when finished:
Because archiving is permanent, confirm both Managed Agents IDs before running the command.
