The agent layer of an iMessage API for AI agents
Retrieval over each customer's knowledge, memory, voice notes, and tool integrations that can't write without permission.
- Client
- MessageBlue
- Through
- Qualkode Technologies
- Role
- Senior Software Engineer
- When
- Jul – Sep 2026
Built as a Senior Software Engineer at Qualkode Technologies, for the Blush group.
- per-app knowledge and per-customer memory across conversations
- RAG + memory
- tool integrations behind a write-confirmation gate
- MCP
- token metering and wallet billing through Stripe
- Prepaid
The situation
MessageBlue gives businesses a real blue-bubble iMessage number that an AI agent can answer. The agent had to know each business's material, remember each customer, take actions in their CRM, and never write to a customer's systems without a person saying yes.
What I built
- Per-app retrieval (RAG) over each business's knowledge, and per-customer memory across conversations.
- Voice note in, voice note out: replies are encoded so they play as native iMessage audio.
- A durable agent task loop that replaced a booking-specific graph.
- Tool integrations for HubSpot, Close and GoHighLevel over MCP, behind a confirmation gate for write actions and a per-integration allow/deny policy enforced in every agent path.
- Slack and Microsoft Teams as channels, alongside iMessage.
- Usage metering and prepaid billing: per-turn token metering, per-app spend caps, and a wallet billed through Stripe. Plus the production environment and CI/CD.
Stack
- TypeScript
- Node.js
- MongoDB Atlas Vector Search
- OpenRouter
- MCP
- Slack & Teams APIs
- Stripe
- GitHub Actions
