Auto Company

Auto Company is a bash-orchestrated loop that turns Claude Code or Codex CLI into a standing “AI company” of 14 expert-persona agents — modeled on real-world thinkers like Jeff Bezos (CEO), Werner Vogels (CTO), Charlie Munger (critic), and DHH (full-stack engineer) — that dynamically form 2-5 person squads each cycle to research demand, make product and pricing decisions, write and deploy code, and market the result, all without a human approving each step.

Each cycle is an independent, stateless CLI invocation. The only thing carried between cycles is memories/consensus.md, a single markdown file the agents read at the start of a cycle and rewrite at the end — a deliberately minimal “consensus memory” instead of a vector database or agent framework. Humans steer the company not through a dashboard button but by hand-editing the “Next Action” line in that file, which the next cycle picks up and acts on. A hardcoded “Forced Convergence Flow” in PROMPT.md pushes cycles through ideation, then validation/pre-mortem, then execution, explicitly forbidding pure discussion so the loop doesn’t stall talking about work instead of shipping it.

The loop itself (scripts/core/auto-loop.sh) handles circuit-breaking on repeated errors, backoff on API rate limits, and log rotation, and can run in the foreground or as a supervised daemon — launchd on macOS, systemd —user on Linux/WSL, with PowerShell wrapper scripts driving the same WSL daemon from native Windows. A local Python dashboard (stdlib-only, no external web framework) exposes live cycle status, cost, and agent activity by shelling out to the same status scripts used from the terminal. Safety guardrails (no repo deletion, no force-push to protected branches, no credential leaks) are written directly into CLAUDE.md as a non-negotiable constitution that every cycle’s LLM call reads before acting, and a library of 30+ reusable skills under .claude/skills/ (deep research, pricing strategy, SEO, security audits, and more) lets any of the 14 agents load a methodology on demand.

What You Get

• A 24/7 bash auto-loop (scripts/core/auto-loop.sh) with circuit-breaker error handling, rate-limit backoff, configurable cycle timeout, and log rotation.
• 14 markdown-defined expert personas under .claude/agents/ (CEO/Bezos, CTO/Vogels, critic/Munger, full-stack/DHH, and 10 more) mapped to documented collaboration workflows for product launches, pricing, and weekly reviews.
• A single-file consensus memory system (memories/consensus.md) that lets a human redirect the AI company by editing its “Next Action” line, with no database or agent framework required.
• A local, dependency-free Python dashboard (dashboard/server.py) plus native daemon installers and monitor scripts for macOS (launchd), Linux/WSL (systemd —user), and Windows (PowerShell wrappers).
• A 30+ skill library under .claude/skills/ covering deep research, pricing strategy, SEO content, security audits, and financial modeling that any agent can load mid-cycle.
• A written safety-guardrail constitution (CLAUDE.md) that hardcodes forbidden actions — no repo/project deletion, no force-push to shared branches, no credential leaks — read by every cycle before it acts.

Common Use Cases

• Running an unattended micro-SaaS factory that ideates, validates, builds, and ships new products in a continuous loop.
• Letting an AI team make day-to-day product, pricing, and go-to-market calls without requiring daily human sign-off.
• Studying multi-agent “expert persona” prompt design and the forced-convergence workflow as a template for stopping agents from looping in discussion instead of shipping.
• Automating a personal or small-team side-project pipeline across whichever OS is available — macOS, native Windows, or Linux/WSL — with a shared local dashboard.
• Accumulating a portfolio of independently AI-shipped projects under projects/, each with its own CI gate and documentation trail.

Under The Hood

Architecture The system is organized around a bash-driven orchestration loop (scripts/core/auto-loop.sh) that invokes an external LLM CLI (Claude Code or Codex CLI) each cycle, passing PROMPT.md and memories/consensus.md as context; the CLI in turn reads CLAUDE.md (charter and guardrails) and .claude/skills/team/SKILL.md (teaming method) to dynamically assemble a squad from the 14 persona-based sub-agents defined as markdown files under .claude/agents/, each mapped to a docs/<role>/ output directory. State persists across cycles only through the single markdown file memories/consensus.md, acting as a hand-off baton between independent, stateless CLI invocations, while .auto-loop-state and .auto-loop.pid track process status and a circuit-breaker/backoff scheme (consecutive-error limits, cooldowns, usage-limit waits) implemented directly in the shell script rather than in a supervising runtime. A Python dashboard exposes a read-only HTTP status view that shells out to the same monitor and status scripts used from the terminal rather than sharing code with the loop, so status-parsing logic is effectively duplicated across bash, PowerShell, and Python; platform-specific process supervision (macOS launchd, Linux/WSL systemd —user, Windows PowerShell wrappers) sits outside the core loop as a thin daemonization layer. If the core abstraction — consensus.md as the sole state-transfer mechanism between stateless CLI calls — were changed, the whole squad-formation and continuity model would break, since there is no database, queue, or structured memory store; this is a monolithic shell script wrapping a stateless LLM subprocess, closer to a cron-orchestrated prompt pipeline than a conventional layered application.

Tech Stack The project is a lightweight orchestration layer around third-party AI CLI engines rather than a typical application stack: package.json declares only project metadata (no listed runtime dependencies), the actual control plane is pure Bash for macOS/Linux/WSL, PowerShell for native Windows control, and a Python 3 standard-library-only dashboard server (http.server.ThreadingHTTPServer, argparse, subprocess — no third-party web framework) paired with a static HTML/CSS/vanilla-JS frontend and no bundler. Process supervision is delegated entirely to OS-native mechanisms — launchd on macOS, systemd —user on Linux/WSL — invoked through wrapper scripts, and the actual compute is outsourced to an external binary, either Anthropic’s Claude Code CLI or OpenAI’s Codex CLI, selected at runtime via an ENGINE environment variable. Continuous integration is a single GitHub Actions workflow scoped narrowly to one generated sub-project under projects/, not to the platform’s own control scripts.

Code Quality Automated test coverage is limited to one unittest-based suite that dynamically loads the dashboard server module and checks its status-string parsing logic against hand-crafted fixture text; there is no test coverage for the bash orchestration scripts, the PowerShell scripts, or the agent persona and skill definitions. Error handling in the shell scripts is defensive at the shell level — strict mode, explicit rate-limit and usage-limit string matching, circuit-breaker counters, stale-PID checks — while the Python dashboard code uses extensive type hints and raises explicit exceptions for unsupported platforms rather than failing silently. Naming is consistent across the codebase, but there is no linter or formatter configuration, and the one CI workflow present gates a specific generated sub-project rather than the platform’s own control-plane code, leaving the core loop and dashboard largely unverified by automation.

What Makes It Unique The genuinely distinctive piece is treating a single markdown file as the entire cross-cycle memory system: rather than a stateful agent runtime or vector store, all context is compressed into one file that’s read and rewritten every cycle, explicitly enabling a human to steer the AI company by hand-editing a “Next Action” line rather than through a dashboard control or API call. This is paired with a hardcoded “Forced Convergence Flow” — an ideation cycle, then a validation/pre-mortem cycle, then an execution cycle, with pure discussion prohibited by the system prompt — meant to stop the agents from looping in deliberation without shipping, and a team-formation mechanism that instantiates a handful of the 14 named expert personas per cycle based on the declared next action. The combination of expert-persona injection, a forced-shipping cadence, and single-file consensus memory is a distinctive prompt-orchestration pattern among AI-agent projects, even though it is not architecturally novel in a systems-engineering sense — the underlying mechanism is a bash loop invoking an external CLI, not a custom execution engine.

Self-Hosting

Licensing Model MIT licensed, per package.json and the README badge — no dedicated LICENSE file was found in the cloned repository, but there are no license-gated code paths anywhere in the codebase.

Self-Hosting Restrictions

• None found — the loop, dashboard, agent personas, and skill library all run locally against your own Claude Code or Codex CLI subscription; there is no server-side or cloud component being restricted.

Enterprise Features

• Not applicable — this is a single-operator local tool with no separate paid tier, license key, or hosted offering described anywhere in the docs.

Cloud vs Self-Hosted

• Not applicable — Auto Company only runs self-hosted, on macOS, Linux/WSL, or Windows via WSL; there is no hosted or cloud version to compare it against.

License Key Required No — the only requirement is an authenticated Claude Code or Codex CLI installation with available model quota; no separate Auto Company license key or account is needed.