Auto Company
Run a persistent, multi-role AI product team on your own computer.
Per-dimension scores and reasoning
The documentation discloses local logs, dashboards, model usage, deployment capability, host/WSL system effects, and redaction of known credentials. CI pins third-party Actions to commit hashes, disables credential persistence, and the README gives detailed contributor and upstream attribution. Deductions are substantial because unattended operation without confirmation is the core design and the product explicitly contemplates danger-full-access or bypassPermissions. The supplied evidence does not show enforced privilege isolation, a complete destination-level data-flow inventory, or general secret detection. Consensus state can be restored after failure, but product changes and external side effects are expressly not rolled back automatically.
The README, CI configuration, and focused tests describe circuit breaking, rate-limit backoff, timeouts, process-tree cleanup, configuration preservation, failure artifacts, and specific error reporting. Dependencies and fail-fast behavior for a missing selected engine are documented. Deductions reflect incomplete dependency version constraints and minor inconsistencies such as describing squads as both 3–5 and 2–5 agents; none of the tests was executed for this static assessment.
The evidence distinguishes macOS, Windows/WSL, engines, languages, and foreground versus daemon use, with concrete environment overrides. Experimental status, platform prerequisites, adapter differences, and the absence of automatic engine fallback are clearly bounded. Next Action steering, cycle stages, and six workflow patterns make routing usable, but trigger selection still depends substantially on natural-language prompts and model judgment, without a rigorous machine-verifiable precedence or conflict scheme, so trigger precision is not full.
The README provides platform navigation, architecture, directory layout, command tables, installation instructions, FAQ material, limitations, contacts, and bilingual documentation links. The complete MIT text agrees with the metadata. Deductions apply because the package name auto-company-clone-win does not fully match the repository product name, version 1.6.0 is present without a supplied changelog or release policy, and although contacts and contributors are identified, the publisher is unverified and no formal maintenance governance or support commitment is shown.
Dashboards, logs, cycle summaries, consensus memory, inspectable project source, and product examples form a practical output chain, while persistent orchestration offers plausible value beyond one-off CLI sessions. Deductions apply because the benefit is supported mainly by repository claims and examples, while every cycle consumes model quota and may create deployment or other external costs. The README itself warns about instability and unexpected output, and no quantified benefit, effective cost ceiling, or comparative user outcome is supplied.
Claims are tied to named files, commands, and documentation, while CI enumerates cross-platform, integration, browser, and product checks, providing useful internal corroboration. The README also distinguishes model-authored titles and summaries from program-collected runtime facts and acknowledges failures, missing data, and human corrections, supporting strong fact/inference separation. Deductions remain because most core implementation files, runtime records, and CI results are absent from the supplied material, so claims such as continuous operation and screenshot provenance cannot be independently confirmed here.
- Treat this as high-risk local automation: do not enable danger-full-access, bypassPermissions, or unattended daemon operation before reviewing its behavior.
- Begin in the foreground with a restricted sandbox, low-quota accounts, no production credentials, and an isolated workspace; continuously inspect projects, docs, logs, and actual network destinations.
- Prompt-based safety rules are not an enforced sandbox. Deployment, marketing, repository writes, and other external effects require additional OS, account, and service-side permission boundaries.
- Consensus rollback is not transactional rollback; code changes, publications, and remote-resource mutations may require manual recovery.
- This assessment uses only the supplied static files and did not execute tests, verify CI status, or inspect omitted core implementation.
What does this agent do, and when should you use it?
Auto Company is a self-hosted multi-agent workflow that continuously coordinates 14 expert roles and more than 30 reusable skills. Its scripts/core/auto-loop.sh process reads PROMPT.md and memories/consensus.md, invokes Claude Code or Codex CLI, and assembles a two-to-five-person squad for the current task. The team can research opportunities, evaluate products, write and test code, deploy projects, and prepare marketing work, with outputs stored under projects/, docs/, and logs/. Cross-cycle state is compressed into memories/consensus.md, whose Next Action can be edited to redirect the next cycle. A local Dashboard presents cycle reports, results, controls, runtime status, usage, budgets, and logs, although it does not track individual-agent activity. It is best suited to adopters prepared to supervise model permissions, quota, host-level risk, and generated work on macOS or Windows with WSL.
On each cycle, auto-loop.sh reads PROMPT.md, CLAUDE.md, .claude/skills/team/SKILL.md, and memories/consensus.md, then starts an independent Claude Code or Codex CLI invocation. Based on the Next Action in consensus.md, it selects two to five roles from the 14-agent roster and performs work such as research, product evaluation, coding, testing, deployment, or marketing. Before the cycle ends, it rewrites memories/consensus.md as the handoff state for the next invocation. Engine-emitted output, cycle results, and available usage records go to logs/; generated applications belong under projects/ and document artifacts under docs/. The loop implements rate-limit waiting, a consecutive-error circuit breaker, timeouts, and consensus recovery after failed cycles, but it does not automatically roll back product-code changes or external side effects. The dashboard/ interface organizes current and historical cycles, runtime controls, status, budgets, usage, and logs, while daemon operation uses launchd on macOS and systemd --user inside Windows/WSL.
- A solo developer wants a local AI team to move repeatedly from product ideation and validation into implementation and launch while retaining review points.
- A product lead wants a structured GO/NO-GO evaluation involving research, executive, inversion, product, technical, and financial roles.
- An engineering team wants feature work routed through interaction design, UI design, full-stack implementation, QA, and DevOps roles.
- A founder wants to steer a persistent automation loop by editing one inspectable Markdown state file instead of operating a separate memory database.
- A macOS or Windows/WSL user wants long-running local automation with cycle history, logs, budget visibility, and circuit-breaker behavior.
What are this agent's strengths and limitations?
- It provides an end-to-end local orchestration loop with dynamic team formation and cross-cycle handoffs, rather than only a collection of role prompts.
- Claude Code and Codex CLI have documented entry points, while macOS and Windows/WSL receive dedicated launchd, systemd --user, and PowerShell control paths.
- The single memories/consensus.md state file is easy to inspect, redirect through Next Action, and restore after a failed cycle.
- The local Dashboard combines cycle reports, results, controls, logs, usage, and budgets while keeping missing, failed, stale, or partial evidence visible.
- Rate-limit backoff, consecutive-error circuit breaking, cycle timeouts, and consensus recovery support unattended experiments.
- The project is explicitly experimental: it runs, but stability is not guaranteed, so continuous autonomous operation still requires human review.
- Every cycle consumes model quota and may cost money; if the selected engine is absent or unavailable, there is no automatic engine fallback.
- High-autonomy operation may rely on Codex danger-full-access or Claude bypassPermissions, allowing system-level actions directly on the host or inside WSL.
- Failure recovery does not automatically reverse product-code changes or external side effects, leaving deployment and publication consequences to the operator.
- Windows operation requires WSL2 and dependencies on both sides of the boundary; a separate native-Linux quick-start commitment is not documented.
- The Dashboard does not expose individual-agent activity, and engine output is not guaranteed to include complete reasoning traces.
How do you install or deploy this agent?
Use macOS, or Windows 10/11 with WSL2 Ubuntu and systemd --user available. Install Python 3.10+, Git, make, Node.js, and either Claude Code or Codex CLI; authenticate the selected CLI and ensure that its account has available model quota. Then run:
git clone https://github.com/MaxMiksa/Auto-Company.git
cd Auto-CompanyFor a first foreground run on macOS or WSL, use make start with the default Claude engine, or ENGINE=codex make start for Codex. For daemon mode on macOS/WSL, use make install, or ENGINE=codex make install. From Windows PowerShell, run .\scripts\windows\start-win.ps1, or .\scripts\windows\start-win.ps1 -Engine codex. Windows users must install and authenticate the chosen CLI inside WSL and also provide Python 3.10+ on Windows for the PowerShell Dashboard entry and local language commands.
How do you use this agent?
Start in the foreground with make start for Claude Code or ENGINE=codex make start for Codex. Use make monitor for live logs, make cycles for cycle summaries, and make dashboard for the local web interface. In daemon mode, use make pause to keep the service stopped and make resume to continue; make stop alone may be followed by an automatic daemon restart. Windows equivalents include .\scripts\windows\status-win.ps1, monitor-win.ps1, dashboard-win.ps1, and stop-win.ps1. To redirect the team, edit the Next Action in memories/consensus.md; the following cycle will read the change. ENGINE, MODEL, LOOP_INTERVAL, CYCLE_TIMEOUT_SECONDS, MAX_CONSECUTIVE_ERRORS, CLAUDE_PERMISSION_MODE, and CODEX_SANDBOX_MODE are among the documented runtime controls. Regularly review docs/, projects/, logs, usage, and budgets rather than treating autonomous output as automatically approved.
How does this agent compare with similar options?
Claude Code is the default engine, while Codex CLI is the documented alternative with its own startup commands. Both must be installed and authenticated separately, and Auto Company does not fail over between them automatically. Cursor and OpenAI-compatible adapters are also named, but they require explicit configuration and may offer different tools and team capabilities, so they are not presented as feature-equivalent replacements.