AI-first dev stack — agent frameworks, hosting, and the tools that ship production agents.

The full stack I use to design, build, and run production AI agents. Three open-source frameworks — Hermes (Nous Research), OpenClaw, and OpenHuman. Self-hosted on Hetzner, Hostinger, and DigitalOcean, wired together with Tailscale, Cloudflare, Supabase, and Postgres. Every choice on this page is something I'm running in production today — not a wishlist.

Looking for the agent-services version of this? See the Hermes & OpenClaw agent stack →

Agent frameworks

The three open-source AI agent frameworks I build production agents on. All self-hosted, all model-agnostic, all yours after the build.

Hermes Agent (Nous Research)

My primary agent framework. Self-improving, persistent memory, scheduled crons, subagent spawning, single gateway across Telegram, Slack, Discord, WhatsApp, Signal, and CLI. Model-agnostic — OpenAI, Anthropic, OpenRouter, Nous Portal, or self-hosted weights.

OpenClaw

Self-hosted, messaging-first agent framework with markdown-based memory (SOUL.md) and a plugin pipeline. The simplest mental model when a Hermes setup is more than the project needs.

OpenHuman

Open-source human-in-the-loop AI agent framework I use when an agent needs explicit approval gates, audit trails, and policy-controlled tool calls before it acts on production systems.

Hindsight (Hermes memory)

Cloud memory layer that pairs with Hermes for cross-session recall — semantic vector store plus structured episodic memory.

Hosting & infrastructure

Where the agents actually run. I default to self-hosted on European VPS providers — Hetzner is my primary, Hostinger is the budget tier, DigitalOcean for clients who already use it. Vercel for the web layer. Tailscale for private networking, ngrok for local-dev tunnels, Cloudflare in front of everything public.

AI code assistants

The four AI coders I rotate between depending on context size, latency, and the kind of work I'm doing.

OpenCode

Primary AI code assistant — terminal-native, Hermes-friendly.

Claude (Code)

Secondary AI assistant — long context, agentic edits, paired with Hermes for scoped tasks.

GitHub Copilot

Tertiary inline completions while typing.

Cursor

AI-native editor for high-context multi-file refactors.

AI memory & MCP servers

Persistent memory and Model Context Protocol servers — what makes an agent stop forgetting and start compounding.

NotebookLM

Deep context and RAG for long documents.

Supermemory

Persistent knowledge layer across sessions and tools.

Ruflo (custom MCP)

My MCP server for agent orchestration — the layer that lets one agent delegate to specialists.

fullstackskills (custom MCP)

Custom 21-agent scaffolding system exposed as an MCP server.

Context7 (MCP)

Live documentation retrieval inside chat — the agent reads the actual docs, not 2024 training data.

Languages & runtimes

What I actually write code in.

Node.js / pnpm

Primary JS runtime and package manager.

Rust / cargo

Systems programming and CLI tooling.

Bun

Fast JS runtime for edge-ready scripts.

Python / uv

AI/ML scripts, fast dependency management.

Java / SDKMAN

Enterprise integrations and JVM tooling.

TypeScript

Strict typing across the entire stack.

Editors

What I open every morning.

VS Code

Primary editor with full extension ecosystem.

Zed

Fast, GPU-accelerated editor for focused sessions.

Neovim

Terminal-native editing and scripting.

Cursor

AI-native editor for high-context sessions.

DevOps & databases

Containers, queues, and stores that keep agents running 24/7.

Podman

Rootless container runtime — primary choice over Docker for agent isolation.

PostgreSQL

OLTP workhorse for agent state, kanban tables, and audit trails.

Supabase

Backend-as-a-service with Postgres, auth, and realtime — the SaaS layer for client work.

Redis

Caching, pub/sub, and session store.

NeonDB

Serverless Postgres for edge workloads.

SOPS / age

Encrypted secrets in Git, decrypted only on the agent host.

Terminal & power tools

The shell tools I'd refuse to live without.

zsh + Starship

Shell and prompt for maximum clarity.

RTK

Custom token killer — 60–90% token savings on dev ops.

lazygit

Terminal UI for git — fast and visual.

eza

Modern ls replacement with icons.

rclone

Cloud storage sync and backup automation.