Run Claude Code like production infrastructure.

Agentic OS in a Box — the harness, routing, self-healing ops, and memory system that run a real business's infrastructure 24/7. Packaged, documented, delivered as a zip. For Claude Code power users who want their agents to remember, route intelligently, heal themselves, and stop asking permission for safe actions.

Get the kit — $99

Why ad-hoc agent use plateaus

You've hit the ceiling. These are the names it has.

Claude Code out of the box is genuinely good at writing code. But at some point, you stop making it better and start making it more. Same session cost, higher bill. Same mistakes you corrected last week. Automations you don't trust to run unwatched.

Agents that forget everything

Every session starts from zero. You re-explain the project, the decisions, the constraints. The agent re-makes mistakes you corrected last Tuesday. Session amnesia is why you can't leave it running overnight.

Surprise API bills

Every request goes to a frontier model because that's the default. No budget ceiling, no cheap-first routing. The bill grows linearly with usage, so you start rationing an assistant you bought to stop rationing your own attention.

Setups that rot after a week

You wire a cron job, it breaks in week two, and you find out in week five. Automation you cannot trust is worse than no automation — you end up supervising the supervisor.

What's inside

Six pieces. One system.

Everything described here is packaged in the kit. Nothing is aspirational. The design is genericized from a real production system running a business's infrastructure 24/7.

Agentic OS Blueprint

4,000-word system design explaining the five subsystems (harness, routing, ops, memory, gates) and how they interlock. Read this before install — it explains why each piece exists.

Hardware Guide

Three tiers from 'enough to run the OS' (one Mac, 16GB) to 'rarely touch an API' (dual-box with GPU). Honest trade-offs at each tier, with real hardware recommendations and cost reality checks.

Harness Pack

Three curated Claude Code skills (autonomous overnight experiments, nightly workflow audit, project tool discovery) plus a session-lifecycle hook and the fail-closed guard patterns you assemble approval gates from.

Ops Pack

Production-grade self-healing: daily health triage script, n8n probe-heal-notify workflow, Telegram alertbot guide, cron/launchd schedules. Silent when healthy, loud when broken.

Routing Pack

Three-tier LiteLLM config: local Ollama → free APIs → paid frontier (budget-capped). 90%+ of requests stay free. Includes escalation doctrine and A/B eval examples.

Memory Pack

File-based persistent memory: index + one-fact-per-file convention, YAML frontmatter format, weekly gardener pass for maintenance. Agents remember decisions across sessions.

Install guide included: afternoon install, single machine primary. Dual-box topology documented in appendix.

Receipts

Not a pitch deck. Real output from the production system.

The system this kit is packaged from runs a real business's infrastructure 24/7. Below is sanitized output from the production system this pack is genericized from — the doctor's silent-when-healthy behavior in action.

[2026-07-19 06:30:14] Doctor: checking 6 containers
[06:30:16] ⚠ ollama           restarted 2h ago (OOM watchdog)

[06:30:18] Log scan: last 24h, 6 containers
[06:30:19]   ollama: 1 error (OOM at 04:22, container restarted by watchdog)
[06:30:20]   All other containers: clean

[06:30:21] Disk: 68% used (218GB / 320GB)

[06:30:22] Triage (local qwen2.5:14b):
  Severity: LOW — system recovered automatically
  Root cause: Ollama OOM likely from untuned max_loaded_models
  Recommendation: cap concurrent models in ollama config
  Action: monitor — if recurs, apply model cap

[06:30:23] Alert sent (1 finding)
[06:30:23] Doctor pass complete.

This is real output. Container names sanitized, disk size rounded. The system that generated this runs every morning at 06:30 and pages only when something is actually broken.

Who this is for

Built for Claude Code power users.

✓ This is for you if:

  • • You use Claude Code daily for real work (not just demos).
  • • You've hit session amnesia, surprise bills, or rotting cron jobs.
  • • You want approval gates without approving every ls.
  • • You're comfortable reading docs, running shell scripts, wiring hooks.

✕ Not for you if:

  • • You've never used Claude Code — this kit assumes familiarity.
  • • You want managed hosting or a SaaS dashboard — this is self-hosted infra you run.
  • • Your team needs hand-holding — support is GitHub issues only, no calls.
  • • You need enterprise SSO, compliance certs, or SLAs — this is a documented kit, not a vendor contract.

Questions before you buy

FAQ

A zip download with six packs: the Blueprint (system design doc), Hardware Guide, and four implementation packs (harness, ops, routing, memory). Working scripts and config examples across the packs, plus step-by-step install docs. Everything you need to deploy the full agentic OS on your own hardware.

Get the kit

Agentic OS in a Box — $99

Instant download. Free updates. 14-day refund window. Support via GitHub issues.

$99
Join waitlist →
  • ✓ One-time payment, no subscription
  • ✓ Free updates forever
  • ✓ 14-day refund via Lemon Squeezy
  • ✓ Support: GitHub issues only