Key takeaways
- The category's autonomy maximalist: 14 persona agents brainstorm, build, deploy, and market products in continuous cycles with zero human intervention by design — "the AI team makes decisions itself, it won't ask you"
- Effectively dormant: the entire commit history spans two days (February 11-12, 2026) and there has been no push since, though stars drifted up to 166 by June 2026
- Architecturally minimal in an interesting way — one `claude -p` call per cycle, plain markdown files as shared consensus memory instead of vector stores, and hard convergence rules to stop agents from planning forever
FAQ
What is auto-company?
auto-company is an open-source (MIT) experiment from GitHub user nicepkg that runs a fully autonomous AI company 24/7 — 14 Claude Code agents with personas modeled on figures like Jeff Bezos, Charlie Munger, and DHH ideate, build, deploy, and promote products with no human in the loop.
How much does auto-company cost?
The software is free and MIT licensed, but the README warns plainly that it "will cost money" — every cycle consumes Claude API credits or Claude Max/Pro subscription quota, and there are no built-in budget caps.
Is auto-company still maintained?
No. As of June 11, 2026 the repository has had no push since February 12, 2026 — roughly four months of silence after a two-day burst of commits. It should be treated as a dormant experiment, not a maintained platform.
How is auto-company different from Paperclip?
Both model AI agents as a company, but Paperclip is a governed orchestration platform where the human acts as the board of directors, while auto-company removes the human entirely — full autonomy, no approvals, and the agents decide what business to run.
Executive Summary
auto-company is the autonomy maximalist of the AI agent company category: a fully autonomous AI company that runs 24/7, in which 14 agents with personas modeled on famous business figures — Jeff Bezos as CEO, Werner Vogels as CTO, Charlie Munger as in-house contrarian, DHH as full-stack developer — brainstorm ideas, write code, deploy products, and try to make money with no human in the loop.[1] Built on Claude Code Agent Teams, the whole system is driven by a single claude -p call per cycle, with plain markdown files serving as shared consensus memory between cycles.[2][3]
Status note: this is effectively a dead project. The repository was created on February 11, 2026, and the last push came on February 12, 2026 — the entire commit history spans two days, and as of June 11, 2026 the repo has been dormant for roughly four months.[1][4] Stars have continued drifting upward (166 as of June 11, 2026, enough to top GitHub's ai-company topic page), but star growth on a dormant repo measures curiosity, not viability.[1][5] What remains is a philosophically interesting artifact: the clearest open-source statement of the zero-human-intervention position, frozen at cycle 28.[4]
| Attribute | Detail |
|---|---|
| Creator | GitHub user nicepkg; primary commit author "2214962083", co-committing with a "claude" account[4] |
| Created | February 11, 2026[1] |
| Last push | February 12, 2026 — dormant since[1] |
| GitHub Stars | 166 (as of June 11, 2026)[1] |
| License | MIT per README; GitHub's API detects no license file[2][1] |
| Funding | None — hobbyist experiment |
Product Overview
You clone the repo, run make start (or make install to register it as a macOS daemon), and walk away. From there the agent company runs in continuous cycles: agents read the shared consensus memory, form ad-hoc teams of 3-5, execute tasks, write results back to memory, and sleep until the next cycle.[2] The creator reported that across the first 12 cycles the system built its own landing page and documentation with only 2 human escalations.[3]
The 14-agent roster is the project's signature. Each agent is a persona of a recognizable figure, chosen for a specific mental model: the Bezos CEO writes PR/FAQs and thinks in flywheels; the Vogels CTO designs for failure; the Munger contrarian runs pre-mortems and reverse thinking; Don Norman, Matías Duarte, and Alan Cooper cover product and interaction design; DHH, James Bach, and Kelsey Hightower handle engineering, QA, and DevOps; Seth Godin, Paul Graham, Aaron Ross, and Patrick Campbell run marketing, operations, sales, and finance; Ben Thompson is the analyst. A library of 30+ specialized skills is available on demand.[2]
Key Capabilities
| Capability | Description |
|---|---|
| 14 persona agents | Strategy, product, engineering, business, and research layers, each modeled on a named figure's mental models[2] |
| 24/7 autonomous cycles | launchd daemon with auto-restart; make start-awake prevents system sleep[2] |
| Auto-convergence | Cycle-based progression (brainstorm → validate → execute) with hard stops — "same action twice = change direction" — to prevent endless planning loops[3] |
| Consensus memory | A single memories/consensus.md markdown file is the cross-cycle state; the creator found plain markdown outperformed vector stores for agent reasoning[2][3] |
| Six standard workflows | Product evaluation, development, release, pricing, reviews, opportunity discovery[2] |
| Safety guardrails | Hardcoded rules in CLAUDE.md block dangerous operations; pause/resume via make pause / make resume[2] |
Technical Architecture
auto-company is deliberately not a framework. The creator positioned it on Hacker News as "not a framework you build on, but a running system" with decision hierarchies baked in — an explicit contrast with AutoGen, CrewAI, and LangGraph.[3] The implementation is mostly Python (77.9%) and shell (12.7%), orchestrated by macOS launchd, with each cycle driven by one headless claude -p invocation against Anthropic's Claude Code CLI.[2][1]
git clone https://github.com/nicepkg/auto-company.git
make start # run in foreground
make install # install as macOS daemon
make status # check on the company
Key Technical Details
| Detail | Value |
|---|---|
| Deployment | Self-hosted, macOS only (Linux/systemd was planned, never shipped)[2] |
| Model(s) | Claude Opus by default; configurable to Sonnet via MODEL=sonnet make start; requires a Claude Max/Pro subscription or API credits[2] |
| Integrations | Claude Code CLI; optional jq, GitHub CLI, Cloudflare Wrangler[2] |
| Open Source | MIT per README; no releases ever published[2] |
Strengths
- The purest articulation of the autonomy thesis — where every competitor adds approval gates, auto-company removes them entirely. The README is blunt: the AI team makes its own decisions and won't ask you.[2] As a philosophical position paper in working code, it is the category's reference point for the no-human-in-the-loop extreme.
- Architecturally honest minimalism — one
claude -pcall per cycle, bash and markdown instead of orchestration frameworks and vector stores. The creator's finding that plain markdown consensus files beat vector stores for agent reasoning is a genuinely useful data point.[3] - Convergence rules that address a real failure mode — the brainstorm → validate → execute progression and "same action twice = change direction" hard stops target the planning-loop problem every multi-agent builder hits.[3]
- The Munger critic worked — per the creator, the contrarian agent's pre-mortems prevented multiple bad decisions, a concrete argument for adversarial roles in agent teams.[3]
- Dogfooded in public — commit history shows the system updating its own promotion materials and documentation through cycle 28, co-authored by a "claude" account.[4]
Cautions
- Dormant for ~4 months — no push since February 12, 2026; the full development history is a two-day burst of 10-12 commits. Nobody is fixing bugs, reviewing PRs, or shipping the planned Linux support.[1][4]
- No human in the loop means no human in the loop — the README disclaims that "AI may build unexpected things" and offers no warranty; guardrails are a hardcoded prompt file, not a governance system.[2]
- Uncapped spend by design — every cycle burns Claude API credits or subscription quota 24/7, with no per-agent budgets or automatic cutoffs. The README's own warning: it "will cost money."[2]
- macOS only, experimental stability — the README describes it as "works but not necessarily stable," and the launchd dependency rules out servers.[2]
- Minimal community validation — the Show HN drew 4 points and 2 comments, one of which was category fatigue: "there are so many of these... what's the difference."[3] 166 stars and 46 forks against 1 watcher and zero releases is curiosity-level traction.[1]
- License ambiguity — the README says MIT, but GitHub's API detects no license file in the repo, which matters if you intend to reuse the code.[2][1]
Pricing & Licensing
| Tier | Price | Includes |
|---|---|---|
| Open source | Free | Full system, all 14 agents, self-hosted on macOS[2] |
Licensing model: MIT per the README, though GitHub detects no LICENSE file — verify before reuse.[2][1]
Hidden costs: the dominant cost is Claude usage — Opus-by-default cycles running 24/7 against API credits or a Claude Max/Pro subscription, with no built-in budget controls.[2]
Competitive Positioning
Direct Competitors
| Competitor | Differentiation |
|---|---|
| Paperclip | The category leader (69K+ stars, actively maintained) takes the opposite governance stance: humans are the board of directors, agents need approval to hire, and per-agent budgets cap spend. auto-company has no board, no approvals, no budgets — and no maintenance. |
| markus | 161 stars on the same ai-company topic page, pitched as an operating system for AI workforces; the closest peer by traction in the topic.[5] |
| auto-co-meta | A 36-star derivative explicitly built around the same 14-agent autonomous-company pattern — evidence the idea outlived the original repo.[5] |
| AutoGen / CrewAI / LangGraph | Frameworks you build agent systems on; auto-company positioned itself as a running system with the decision hierarchy already baked in.[3] |
When to Choose auto-company Over Alternatives
- You want to study the zero-human-intervention design pattern — persona agents, markdown consensus memory, convergence rules — and a frozen, readable codebase is a feature, not a bug
- You run macOS, hold a Claude Max subscription, and want a weekend experiment in letting an agent company loose
- You are prototyping your own autonomous-company system and want MIT-licensed (per README) prior art to dissect
For anything you intend to keep running, choose an actively maintained alternative.
Ideal Customer Profile
Best fit:
- Researchers and writers studying autonomous multi-agent design patterns
- Claude Code power users who want a contained, disposable experiment
- Builders mining a small, legible codebase for ideas (consensus memory, critic agents, convergence rules)
Poor fit:
- Anyone who needs the software to be maintained, patched, or supported
- Teams that require human approval gates, audit trails, or spend controls
- Linux/server deployments — macOS launchd is a hard dependency[2]
- Anyone expecting the "make money" part of the pitch to be a proven outcome rather than an aspiration
Viability Assessment
| Dimension | Assessment |
|---|---|
| Financial Health | N/A — unfunded hobbyist experiment with no commercial entity |
| Market Position | Top repo on GitHub's small ai-company topic (28 repos), but at 166 stars the whole niche is early[5][1] |
| Innovation Pace | Stopped. All development occurred February 11-12, 2026; no releases were ever published[4][2] |
| Community/Ecosystem | 46 forks, 1 watcher, 5 open issues, a 2-comment Show HN — interest without participation[1][3] |
| Long-term Outlook | Dormant. Its ideas are already propagating through derivatives like auto-co-meta, which may be its real legacy[5] |
auto-company should be evaluated as an artifact, not a product. It compressed a complete worldview — full autonomy, persona-based agent design, radical infrastructural minimalism — into two days of commits, posted the results, and stopped. The 30-star drift since spring shows the idea still attracts attention; the four months of silence show no one, including its creator, is carrying it forward in this repo.
Bottom Line
auto-company is the most philosophically committed project in the agent-company category and also the most clearly abandoned. As a working argument that an AI company needs no human in the loop — just personas, markdown memory, convergence rules, and a Claude subscription — it is worth an afternoon of any agent builder's time. As infrastructure, it is a two-day-old codebase that has been unmaintained since February 2026, runs only on macOS, and burns Opus tokens around the clock with no budget controls.
Recommended for: studying autonomous multi-agent design patterns, mining prior art, or running a contained weekend experiment.
Not recommended for: any sustained or production use — there is no maintainer, no releases, no support, and no governance model beyond a prompt file.
Outlook: dormant and likely to stay that way. The interesting question is no longer whether auto-company succeeds but whether its maximalist pattern — already echoed by derivatives on the same topic page — gets absorbed by maintained platforms like Paperclip as an optional mode rather than a founding principle.
Research by Ry Walker Research • methodology