Mission Control

## Inspiration

Every day I spend 2-3 hours processing digital exhaust—reading meeting transcripts for the 3 action items, scanning 50 emails to find the 2 that matter, reviewing PRs to figure out which one actually needs my eyes. AI agents can automate workflows, but there's a missing layer: **where does the human make decisions?**

Teams are building agents and workflows with Retool, connecting them to Claude, wiring up webhooks—but the output scatters across Slack, email, dashboards, and databases. The human-in-the-loop becomes the human-lost-in-the-chaos.

Mission Control is the orchestration layer that's been missing.

## What it does

Mission Control receives pre-processed work items from AI agents and displays them in **one unified queue**, sorted by priority. 

- **Meeting Agent** (via FireCrawl) extracts decisions and action items from transcripts
- **Email Agent** (via Gmail) classifies and triages incoming messages  
- **GitHub Agent** analyzes PRs for risk and routes them appropriately

All outputs flow into a single task queue. The user sees: "You have 7 things that need your brain. Everything else is handled."

The goal: **15 minutes to clear your queue.**

## How we built it

**Retool Workflows** handle the agent orchestration—webhooks trigger Claude-powered classification and extraction, then POST structured tasks to our API.

**Replit** hosts the React dashboard and Node/Postgres backend. One unified `tasks` table stores everything. Source data (full transcripts, email bodies, PR details) lives in separate tables, loaded on-demand.

**Anthropic Claude** powers the intelligence—classifying emails, extracting meeting action items, assessing PR risk.

**ElevenLabs** generates a daily 5-minute audio briefing so you can hear your day while drinking coffee.

The architecture separates concerns cleanly: Replit owns the data and UI, Retool owns the AI orchestration.

## Challenges we ran into

**The "one queue" insight took time.** We initially built separate tabs for emails, PRs, and meetings. It felt organized but was wrong—users don't care about source, they care about priority. Collapsing everything into one prioritized queue was the unlock.

**Prompt engineering for consistent JSON.** Getting Claude to return parseable JSON reliably required explicit instructions and fallback parsing. "Return ONLY valid JSON" became our mantra.

**Webhook reliability.** Coordinating Replit polling Gmail, posting to Retool, and Retool posting back to Replit required careful error handling and idempotency checks.

## What we learned

Agents without a control plane are chaos. The orchestration layer—visibility, governance, human-in-the-loop—is just as important as the AI itself. 

This is what we call **context engineering**: not just building agents, but designing the information flow so the right context reaches the right decision-maker at the right time.

## What's next

- **More input sources**: Slack, Linear, calendar
- **Action execution**: Approve a PR from the dashboard and Retool merges it
- **Learning from decisions**: Track which AI suggestions get accepted/rejected to improve over time
- **Team support**: Multiple users, delegation, shared queues

Built With

• anthropic
• elevenlabs
• replit
• retool