Going All In on REST Connectors at Block: Vibe coding with Goose
summary
Block converts gRPC protos to JSON and uses Goose AI agent plus Apollo Connectors and an Apollo MCP server to auto-generate REST connector subgraphs; a proof of concept puts one connector into staging, a hack week with 8–12 teams finishes multiple connectors in 4–5 days, and platform-wide totals target 700 teams / ~3,000 services. Workflows use reusable recipes and sub-recipes, a Flock multi-agent pipeline with prepare/process/review/merge queues, MCP-driven tests and UI embedding; generated schemas are mostly correct but live demos prove flaky.
AI-written summary · about
chapters (11)
- 0:00REST connectors and AI journey
- 1:48Block's investment in REST connectors
- 4:37AI agents and Goose for code
- 7:03MCP servers for AI capabilities
- 9:12Goose recipes for AI workflows
- 12:03Sub-recipes and plan-execute-reflect
- 14:28Multi-agent architecture with Flock
- 22:57GraphQL APIs with AI workflows
- 25:18Generative UI with AI and MCP UI
- 33:14Scaling AI workflows and future
- 34:18Q&A
related talks
26mOrchestrating APIs for LLMs - Agentic Workflows for GraphQL
30mSchema Design Patterns: Leveraging Existing REST APIs for Rapid GraphQL
01h 23mWorkshop: Apollo Router & MCP: A Modern Agentic Development Approach
30mFrom Experiment to Enterprise: How Block Operationalized MCP at Scale
26mAsk, Don't Fetch: Why GraphQL + MCP Is the Native Language of AI Agents
45mPowering AI Agents with Temporal and Kgoose
26mBuilding APIs for Developers & Machines
34mFrom "Vibes" to Production: Eliminating RAG Data Silos with MCP