What You're Looking At

An AI coding assistant installs VibeSQL from npm, creates a 10-table agent mail schema, registers agents, sends messages between them, manages kanban boards, runs queries, and tears everything down. No faking. No local paths. Everything from the public npm registry, exactly how a real user would do it.

It took 68 takes to get here.

Why 68 Takes

Because shipping real developer tools is hard, and we refused to cheat.

Every take was a real npm install from the public registry. Every MCP tool call went through the actual protocol. When something broke, we fixed the root cause and published a new version — we didn't patch around it in the script.

Here's what we fixed along the way:

vibesql-mcp (MCP Server) — 4 releases in one day:

Version	What broke	What we fixed
1.0.0	Health check killed the process on startup (MCP starts before the database)	—
1.0.1	Removed the blocking health check	Published with stale build artifacts
1.0.2	No visibility into what the MCP server was doing	Added stderr logging
1.0.3	`list_tables` returned nulls, `query` crashed on every call	Rewrote the API response transformer — vibesql-micro returns rows as objects, MCP expected arrays

vibesql-micro (Database) — 4 releases in one day:

Version	What it added
1.0.1	Windows PostgreSQL `/share` directory fix
1.0.2	Real-time startup diagnostics instead of "timeout after 30s"
1.0.3	Parameterized queries (`$1`, `$2`) for SQL injection prevention
1.0.4	Graceful shutdown via `POST /v1/shutdown` and `vibe stop` command

Other discoveries:

MCP config goes in .mcp.json at the project root, not .claude/settings.json
Windows needs cmd /c wrapper for npx in MCP configs
npx caches aggressively — pin the version (vibesql-mcp@1.0.3) or it serves stale code
Pre-approve MCP tools in settings.local.json or the demo hits permission prompts on every call
A global /vibe-sql skill was teaching the AI to use curl instead of MCP tools — had to delete it

The Schema

The demo builds a full agent mail system — 10 tables, 14 indexes:

settings — key-value configuration
agents — agent profiles with role definitions, expertise, and identity markdown
teams — agent team groupings
team_members — many-to-many agent-team relationships
projects — project tracking
kanban_boards — visual boards linked to projects
kanban_columns — board columns (Backlog, In Progress, Done)
kanban_cards — cards with agent assignments
messages — threaded messages with importance levels
inbox — per-agent delivery with read/archive tracking

This schema mirrors the production vibe_agents collection from the PayEz VibeDB — translated from JSONB documents to relational PostgreSQL tables.

What MCP Actually Does

MCP (Model Context Protocol) is a JSON-RPC protocol over stdio. The AI tool spawns the MCP server as a child process and calls tools through structured messages. No HTTP, no REST, no curl.

The vibesql-mcp package exposes 9 tools:

query          — Execute any SQL (DDL + DML, unrestricted)
list_tables    — List all tables
describe_table — Column schema
table_data     — Browse rows with pagination
create_table   — DDL helper (validates, prevents multi-statement injection)
insert_row     — Insert from JSON key-value pairs
help           — VibeSQL topic help
help_products  — Product family overview
help_architecture — Architecture patterns

The AI uses these tools natively — no curl, no HTTP clients, no bash commands. It calls query with SQL and gets back formatted results.

Try It Yourself

npm install -g vibesql-micro
npx vibesql-micro serve

Add .mcp.json to your project:

{
  "mcpServers": {
    "vibesql": {
      "command": "npx",
      "args": ["--yes", "vibesql-mcp@1.0.3"]
    }
  }
}

On Windows, use cmd /c:

{
  "mcpServers": {
    "vibesql": {
      "command": "cmd",
      "args": ["/c", "npx", "--yes", "vibesql-mcp@1.0.3"]
    }
  }
}

Open Claude Code in that directory. The MCP tools appear automatically. Ask it to create tables, insert data, run queries — it uses the tools directly.

What's Next

vibesql-mail — a dedicated Rust API for agent-to-agent messaging (the schema from this demo, but as a standalone service with its own MCP server)
GPT Store — OpenAPI spec for vibesql-micro so Custom GPTs can query your database
Gemini — function declarations for the same tool set

The video is up. The packages are on npm. The code is Apache-2.0.

68 takes. Zero shortcuts.