The minimum-viable AI stack for a one-person shop. Claude Code as the coding agent, a vector DB for product-knowledge memory, an observability tool to catch model regressions before customers do, and a sandbox so AI-generated code never touches your laptop. Picked for cost: every tool here has a generous free tier and a clear paid step-up when revenue justifies it.
What's in this stack
- Claude CodeAnthropic's terminal-native coding agent with MCP-native tool use and skills system.npm_package
- AnthropicClaude model family — the default Claude Code provider. Sonnet, Opus, Haiku.env_key
- Filesystem MCP ServerOfficial MCP reference server for safe local filesystem access (read/write within allowed paths).mcp_npm
- GitHub MCP ServerOfficial MCP server for GitHub — repos, issues, PRs, search, and code via the GitHub API.mcp_npm
- pgvectorVector similarity search as a PostgreSQL extension. The 'boring tech' pick for RAG.external_link
- LangfuseOSS, self-hostable LLM observability — traces, evals, prompts, and datasets.pip_package
- E2BSecure cloud sandboxes for AI-generated code — Python, Node, and full Linux environments.pip_package
How to use it
Click any entity above for its install instructions. Most are one-line installs; the MCP servers are JSON-config additions to your Claude Code config file.
The free Arcade tier lists these entities for browsing + per-entity install. Arcade Pro adds 1-click bulk-install: with one button in the desktop app, the entire stack lands configured and ready.