Duolingo AI Slackbot – Aaron Wang
Speaker: Aaron Wang, Software Engineer (DevXAI), Duolingo
Journey Timeline
| Date | Milestone |
|---|---|
| Nov 2024 | MCP introduced, engineers start experimenting |
| May 2025 | Centralized "MCP store" with setup instructions |
| Aug 2025 | MCP standardization efforts begin |
| Sep 2025 | AI Slackbot launches |
| Apr 2026 | 250+ weekly users (30% of company), 80% approval rate |
Architecture
- Claude Code Agent SDK — AI brain
- Slack Bot API — messaging interface
- 50+ MCP servers connected
- Tools: AWS, BigQuery, internal services, external APIs
Standardization Strategy
| Server Type | Approach |
|---|---|
| First-party (Linear, etc.) | Connect directly from AI client |
| Open-source | Fork, add auth, host internally via HTTP |
| Shared credentials (Jenkins) | Shared service token on internal server |
| Individual credentials (Google, Slack) | Run on internal service VM |
| Internal services | Python library to convert to MCP, host internally |
| IAM-controlled | Use add tools, not MCP |
| Local-only (Playwright) | Support via stdio |
Security Principles
- Bot is not a way to bypass permissions — different access layers by group
- No write operations without human approval — only Slack responses autonomous
- Sandboxed environment — no access to other machine resources
- No side channel between users — messages sandboxed per-person
Key Features
- Respond in channels + DMs
- Group verification for write operations
- System prompts for internal info sources
- Channel-specific documentation
- Power skills and starter agents
- Feedback collection + regression tests
Open Source
Repository: github.com/duolingo/slack-ai-agents
Key Quote
"Even a single click to copy a config is still too much friction. We wanted to provision MCP servers to people automatically."