Skip to content
/
Tier A A11 Governance

Agentic Coding Flywheel Setup (ACFS)

flywheel-sdd · Dicklesworthstone/agentic_coding_flywheel_setup · ★ 1.5k · last commit 2026-05-24

Take a complete beginner from laptop to three AI coding agents running on a VPS in 30 minutes via an idempotent manifest-driven Bash installer.

Best whenThe bottleneck to AI-assisted development for most people is environment setup, not methodology — fix the infrastructure first.
Skip ifDeleting files without explicit permission, Running destructive git commands without user authorization
vs seeds
agent-osinstalls Claude Code commands/standards into an existing project whil…
Primitive shape
No installable primitives
00

Summary

The Agentic Coding Flywheel (ACFS) — Summary

The Agentic Coding Flywheel Setup (ACFS) is a Bash bootstrapping system for Ubuntu VPS servers that installs a complete multi-agent AI development environment in ~30 minutes via a single curl | bash command. It is not a prompt framework — it is an infrastructure provisioning tool. ACFS installs three AI coding agents (Claude Code, Codex CLI, Gemini CLI), 30+ developer tools (language runtimes, cloud CLIs, agent coordination tools), configures tmux for multi-session management, and sets up the "Dicklesworthstone stack" of agentic utilities. The framework has a companion wizard website at agent-flywheel.com that guides beginners through renting a VPS, connecting via SSH, and running the installer. A manifest-driven architecture (acfs.manifest.yaml) serves as the single source of truth for all installed modules.

Compared to seeds, ACFS is unlike all 11 seeds — it is closer to agent-os (both are bash-bundle distributions that set up development infrastructure) but where agent-os installs Claude Code commands/standards into a project, ACFS provisions a complete server environment. It is infrastructure-as-code for agentic development, not a prompt or spec framework.

01

Overview

Agentic Coding Flywheel Setup — Overview

Origin

Created by Dicklesworthstone (Jeffrey Emanuel). Version 0.7.0. MIT + OpenAI/Anthropic Rider license. 1,488 stars, 2 contributors. Shell language. Active: last push May 2026.

The companion website is at agent-flywheel.com — a 13-step wizard (Next.js 16) for complete beginners.

Philosophy

"From zero to fully-configured agentic coding VPS in 30 minutes."

"Zero to Hero: Takes complete beginners from 'I have a laptop' to 'I have Claude/Codex/Gemini agents writing code for me on a VPS'"

"Vibe Mode: Pre-configured for maximum velocity — passwordless sudo, dangerous agent flags enabled, optimized shell environment"

The "flywheel" metaphor: once the environment is set up, agents can build on each other's work in a fast iteration loop. The setup investment pays off through accelerated subsequent development cycles.

Key Design Principles

From AGENTS.md:

"RULE NUMBER 1: NO FILE DELETION. YOU ARE NEVER ALLOWED TO DELETE A FILE WITHOUT EXPRESS PERMISSION."

"The installer is idempotent — if interrupted, simply re-run it. It will automatically resume from the last completed phase without prompts."

Manifest Architecture

From README:

"ACFS is built around a single source of truth: the manifest file. Everything else — the installer scripts, doctor checks, website content — derives from this central definition."

The acfs.manifest.yaml defines every tool with: id, description, category, phase, dependencies, installed_check commands, install commands, and verify commands.

Vibe Mode

The --mode vibe flag enables:

  • Passwordless sudo for the target user
  • Dangerous agent flags enabled for Claude Code
  • Optimized shell environment (zsh + oh-my-zsh + powerlevel10k)
02

Architecture

Agentic Coding Flywheel Setup — Architecture

Distribution

  • Type: Bash script bundle (curl | bash installer)
  • Install: curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/agentic_coding_flywheel_setup/main/install.sh?$(date +%s)" | bash -s -- --yes --mode vibe
  • License: MIT + OpenAI/Anthropic Rider
  • Language: Bash (installer), Next.js 16 (wizard website), TypeScript (Bun/Next.js web)
  • Version: 0.7.0
  • Platform: Ubuntu 25.10 (auto-upgrades from 22.04+)

Directory Structure

install.sh              ← Main idempotent installer (Bash, ~1000+ lines)
acfs.manifest.yaml      ← Single source of truth: all tools, phases, verify
checksums.yaml          ← Pinned checksums for reproducible installs
VERSION                 ← Version file

acfs/
  claude/
    settings.json       ← Base Claude Code settings (hooks registered separately)
  tmux/
    tmux.conf           ← tmux configuration
  zsh/                  ← zsh/oh-my-zsh config
  onboard/              ← Onboarding scripts
  templates/            ← Project templates
  docs/                 ← Documentation

apps/
  web/                  ← Wizard website (Next.js 16, deployed to Vercel)

packages/               ← Shared packages (Bun monorepo)

scripts/
  tests/                ← Installer test suite

Manifest-Driven Phases

The installer runs in sequential phases:

  1. Base system — apt packages, curl, git, jq
  2. User normalization — target user, passwordless sudo, SSH keys
  3. Filesystem setup — workspace directories, ACFS config
  4. Language runtimes — bun, uv/Python, Rust, Go
  5. Shell environment — zsh, oh-my-zsh, powerlevel10k, history
  6. AI coding agents — Claude Code, Codex CLI, Gemini CLI
  7. Agent coordination — NTM, MCP Agent Mail, SLB
  8. Cloud CLIs — Vault, Wrangler, Supabase, Vercel CLI
  9. Additional developer tools — 20+ tools

Idempotency

The installer checkpoints each completed phase. Re-runs skip completed phases. --force-reinstall starts fresh. --resume (default) continues from checkpoint.

Target Environment

Fresh Ubuntu VPS (OVH, Contabo, or similar). Not for macOS or Windows (the wizard website is for the local machine that SSHes into the VPS).

03

Components

Agentic Coding Flywheel Setup — Components

ACFS ships no Claude Code commands, skills, hooks, agents, or spec templates. Its components are infrastructure modules.

Installed Tools (from manifest)

AI Coding Agents

  • Claude Code (with --dangerously-skip-permissions in vibe mode)
  • OpenAI Codex CLI
  • Gemini CLI

Agent Coordination Stack ("Dicklesworthstone Stack")

  • NTM (Node Task Manager) — multi-agent task coordination
  • MCP Agent Mail — agent-to-agent communication
  • SLB (Simple Load Balancer for agents)
  • Additional Dicklesworthstone utilities (10 tools)

Language Runtimes

  • Bun (JavaScript/TypeScript)
  • uv + Python
  • Rust + cargo
  • Go

Shell Environment

  • zsh + oh-my-zsh + powerlevel10k
  • History optimization
  • Shell tracing (ACFS_DEBUG=true → set -x)

Cloud CLIs

  • HashiCorp Vault
  • Cloudflare Wrangler
  • Supabase CLI
  • Vercel CLI

Developer Tools

  • tmux (with custom .tmux.conf)
  • PostgreSQL 18 (optional)
  • Docker (optional)
  • Git with configured aliases
  • shellcheck, jq, and more

CLI Flags (install.sh)

Flag Purpose
--yes Skip all prompts, use defaults
--mode vibe Enable passwordless sudo, dangerous agent flags
--dry-run Print what would be done without changing system
--list-modules List available modules
--print-plan Print execution plan
--only <module> Run only specific module
--skip <module> Skip specific module
--ref <ref> Pin to specific git ref
--offline-pack <dir> Use offline package cache
--force-reinstall Start fresh
--resume Resume from checkpoint (default)

Website (agent-flywheel.com)

13-step wizard for beginners:

  1. Install terminal on local machine
  2. Generate SSH keys
  3. Rent a VPS
  4. Connect via SSH
  5. Run the installer
  6. Reconnect with SSH key
  7. Start coding with agents
05

Prompts

Agentic Coding Flywheel Setup — Prompt Excerpts

ACFS ships no prompt files for end users. The AGENTS.md at the repo root is instructions for Claude Code working on ACFS itself.

Excerpt 1: AGENTS.md — Safety Rules for Claude Working on ACFS

Source: AGENTS.md

## RULE NUMBER 1: NO FILE DELETION

**YOU ARE NEVER ALLOWED TO DELETE A FILE WITHOUT EXPRESS PERMISSION.** Even a new file that
you yourself created, such as a test code file. You have a horrible track record of deleting
critically important files or otherwise throwing away tons of expensive work. As a result, you
have permanently lost any and all rights to determine that a file or folder should be deleted.

**YOU MUST ALWAYS ASK AND RECEIVE CLEAR, WRITTEN PERMISSION BEFORE EVER DELETING A FILE OR
FOLDER OF ANY KIND.**

---

## Irreversible Git & Filesystem Actions — DO NOT EVER BREAK GLASS

1. **Absolutely forbidden commands:** `git reset --hard`, `git clean -fd`, `rm -rf`, or any
   command that can delete or overwrite code/data must never be run unless the user explicitly
   provides the exact command and states, in the same message, that they understand and want
   the irreversible consequences.

Prompting technique: Tone escalation — "YOU HAVE A HORRIBLE TRACK RECORD" and "permanently lost any and all rights" are unusually strong negative framing. This is an extreme version of the Iron Law pattern, written with evident frustration from real incidents.

Excerpt 2: AGENTS.md — Destructive Command Protocol

4. **Mandatory explicit plan:** Even after explicit user authorization, restate the command
   verbatim, list exactly what will be affected, and wait for a confirmation that your
   understanding is correct. Only then may you execute it — if anything remains ambiguous,
   refuse and escalate.
5. **Document the confirmation:** When running any approved destructive command, record (in the
   session notes / final response) the exact user text that authorized it, the command actually
   run, and the execution time. If that record is absent, the operation did not happen.

Prompting technique: Audit trail mandate — requiring the agent to document the authorization text + command + timestamp for any destructive action. Creates an accountability chain beyond what any other framework in this batch specifies.

09

Uniqueness

Agentic Coding Flywheel Setup — Uniqueness

Differs From Seeds

ACFS is unlike all 11 seeds — it is an infrastructure provisioning tool, not a prompt or spec framework. The closest seed is agent-os (both are bash bundles that set up AI coding agent environments), but agent-os installs Claude Code methodology (commands, standards) into an existing project, while ACFS provisions a complete Ubuntu VPS server from scratch. ACFS is one abstraction level below: it creates the environment that agent-os and other frameworks run inside.

Positioning

  • "Zero to Hero" for complete beginners
  • VPS-first: the framework assumes the user needs a server, not just a local dev environment
  • "Vibe Mode" as a first-class configuration — maximum velocity over security caution
  • Infrastructure as code: manifest-driven, reproducible, idempotent

Notable Patterns

  1. AGENTS.md tone escalation: The most confrontational agent instruction file in this batch — "YOU HAVE A HORRIBLE TRACK RECORD" — suggesting real incidents drove the rule writing
  2. Manifest single source of truth: acfs.manifest.yaml drives installer scripts, doctor checks, AND website content — unusual tight coupling
  3. Offline pack support: --offline-pack <dir> enables air-gapped installs
  4. MIT + OpenAI/Anthropic Rider license: An unusual dual-license that adds terms for AI companies

Observable Failure Modes

  1. Ubuntu 25.10 specific: Only officially supports Ubuntu 25.10 (auto-upgrades from 22.04+ but no macOS/Windows support)
  2. Vibe mode security: --mode vibe enables passwordless sudo and dangerous agent flags — inappropriate for multi-user or shared environments
  3. Not a methodology: ACFS gives you tools but no guidance on how to use them effectively
  4. Scope mismatch: Researchers looking for a spec/prompt framework will not find one here — ACFS is infrastructure

Cross-References

  • The "Dicklesworthstone stack" (NTM, MCP Agent Mail, SLB) are separate tools by the same author
  • agent-flywheel.com wizard suggests ACFS is designed to convert people from "I heard about AI coding" to "I'm running agents" — the earliest onboarding ramp in this batch
04

Workflow

Agentic Coding Flywheel Setup — Workflow

ACFS has no development workflow. It is a one-time setup tool. The "workflow" is installation.

Installation Workflow

Beginner with laptop
  → Visit agent-flywheel.com (13-step wizard)
  → Get VPS (OVH, Contabo)
  → SSH into fresh Ubuntu VPS
  → Run: curl -fsSL "<installer>" | bash -s -- --yes --mode vibe
  → ~30 minutes
  → Fully configured environment with 3 AI coding agents ready

Idempotent Phase Execution

Each manifest module has:

  1. installed_check — command to verify if already installed
  2. install — installation commands
  3. verify — post-install verification commands

The installer runs each module sequentially, skipping already-installed ones.

Post-Install "Flywheel" Concept

After installation, the user:

  1. SSHes into VPS
  2. Opens tmux sessions
  3. Runs Claude Code / Codex / Gemini in separate panes
  4. Agents build features in parallel
  5. NTM/MCP Agent Mail coordinates between agents

The "flywheel" effect: each agent's output becomes input for the next agent session, building momentum over time.

Approval Gates

None during installation (with --yes). The wizard website has 7 human steps. Post-install agent usage is entirely up to the user.

06

Memory Context

Agentic Coding Flywheel Setup — Memory & Context

No framework memory or context management. ACFS installs the environment; what agents do after install is entirely determined by the user and the agent tools.

Installed Memory Infrastructure

ACFS installs tools that other frameworks use for memory:

  • Claude Code (which reads CLAUDE.md / AGENTS.md in projects)
  • NTM (Node Task Manager) — may provide cross-session task state
  • The base Claude Code settings.json is deployed at acfs/claude/settings.json (currently minimal, hooks registered separately via dcg install)

No Spec Format or Context Injection

ACFS ships no spec templates, CLAUDE.md files, or context injection for user projects. After install, the user starts from a clean slate with powerful tools.

07

Orchestration

Agentic Coding Flywheel Setup — Orchestration

ACFS provides the infrastructure for multi-agent coordination but does not define an orchestration protocol.

Multi-Agent Infrastructure Installed

  • tmux: Multiple terminal sessions for running agents in parallel panes
  • NTM (Node Task Manager): Agent task coordination tool (from Dicklesworthstone stack)
  • MCP Agent Mail: Agent-to-agent communication
  • SLB (Simple Load Balancer): Agent load distribution

No Programmatic Orchestration

ACFS does not define how agents should coordinate. It installs the tools that enable coordination; the specific protocol is up to the user and the agents.

Execution Mode (of ACFS itself)

One-shot installation. Sequential phase execution.

Multi-Model

Yes — installs three models (Claude Code, Codex, Gemini) but no programmatic routing between them.

08

Ui Cli Surface

Agentic Coding Flywheel Setup — UI / CLI Surface

CLI Binary

The installer (install.sh) has a rich CLI surface:

curl -fsSL "<url>" | bash -s -- [OPTIONS]

Options: --yes, --mode vibe, --dry-run, --print, --list-modules, --print-plan, --only <module>, --skip <module>, --ref <ref>, --offline-pack <dir>, --force-reinstall, --resume, --strict, --interactive

Web UI

Yes — agent-flywheel.com:

  • 13-step wizard for complete beginners
  • Built with Next.js 16 (App Router), React 19, Tailwind CSS 4, Bun
  • Deployed on Vercel + Cloudflare
  • Interactive steps guide SSH key generation, VPS setup, and installer execution

Local UI

None. After install, the user works in their terminal with tmux for session management.

Observability

  • --dry-run: prints what would be done
  • --print-plan: shows execution plan
  • --list-modules: shows all installable modules
  • Health checks: each module has a verify block that runs post-install
  • ACFS_DEBUG=true bash install.sh: enables shell tracing (set -x)
  • checksums.yaml: reproducibility via pinned checksums

Related frameworks

same archetype · same primary tool · same memory type

OpenHarness ★ 13k
A11 Governance

Open-source Python agent runtime providing complete harness infrastructure: tools, memory, governance, swarm coordination, and…

Trae Agent ★ 12k
A11 Governance

Research-friendly open-source CLI coding agent by ByteDance, designed for academic ablation studies and modular LLM provider…

Sweep AI ★ 7.7k
A11 Governance

Autonomous GitHub bot that converts issues to pull requests using a sequential multi-agent pipeline.

Agent Governance Toolkit (microsoft) ★ 2.3k
A11 Governance

Enterprise-grade AI agent governance: YAML policy enforcement, 12-vector prompt injection defense, zero-trust identity,…

TDD Guard ★ 2.1k
A11 Governance

Mechanically enforces the Red-Green-Refactor TDD cycle by blocking file writes that violate TDD principles via a PreToolUse hook…

leash (strongdm) ★ 565
A11 Governance

Wraps AI coding agents in containers with eBPF-enforced Cedar policies, making policy violations (unauthorized file access,…