Company-as-Code

A collaborative design project between Claude (Anthropic) and Codex (OpenAI), with a human proxy orchestrating the conversation.

What This Is

Designing a new product from scratch: a Company-as-Code operating system — where a company's state, doctrine, decisions, and agent contracts are modeled as a typed, versioned, testable system.

How This Works

1. Claude and Codex exchange ideas via markdown files in discussion/
2. The human proxy carries context between agents
3. Research and reference material lives in references/
4. Product design emerges in design/
5. Everything is versioned in git — the conversation IS the artifact

Structure

company-as-code/
├── README.md                         # This file
├── references/                       # Research & analysis
│   ├── openclaw.md                  # OpenClaw (agent runtime)
│   ├── mission-controls.md          # Three Mission Control variants compared
│   ├── onecontext.md                # OneContext (unified agent context)
│   ├── bhanu-architecture.md        # Bhanu's 10-agent squad (reference impl)
│   └── company-as-code-concept.md   # The original concept definition
├── discussion/                       # Agent-to-agent conversation
│   ├── PROTOCOL.md                  # How the discussion works
│   └── 001-claude-synthesis.md      # Opening move
├── design/                           # Emergent product design
│   └── (TBD — emerges from discussion)
└── repos.md                          # External repo references

Setup

• pnpm install && pnpm build
• Install OpenClaw and start the gateway (see the OpenClaw repo).
• Add agent IDs to ~/.openclaw/openclaw.json under agents.list:

{
  "agents": {
    "list": [
      { "id": "codex", "default": true },
      { "id": "claude" },
      { "id": "ops" },
      { "id": "example" }
    ]
  }
}

• Export the gateway token: export OPENCLAW_GATEWAY_TOKEN=...
• Run node dist/cli.js sync to verify.

Install

• Global install: npm i -g company-as-code
• One-off run: npx company-as-code --help

Requires Node 22+.

Commands

• node dist/cli.js init — scaffold a repo with Company-as-Code structure
• node dist/cli.js check — run organizational tests
• node dist/cli.js check --watch — re-run checks on file changes
• node dist/cli.js validate — structural validation of core repo files
• node dist/cli.js diff — compare repo changes between git refs
• node dist/cli.js diff --semantic — summarize YAML changes by item id
• node dist/cli.js normalize — backfill discussion frontmatter
• node dist/cli.js view — render objectives, risks, pipeline, and work queue
• node dist/cli.js status — show handoff log, work queue, and check summary
• node dist/cli.js log — show handoff history (pretty table or JSON)
• node dist/cli.js sync — compile/sync agent contracts (--compile-only for offline)
• node dist/cli.js handoff — generate a relay prompt (--send to deliver)
• node dist/cli.js receive — create a response stub (--format json for agent runtimes)
• node dist/cli.js simulate — dry-run handoff/receive and write reports/simulation.json
• node dist/cli.js doctor — diagnose repo health (validate + check + compile + env)

Handoff (Level 3)

• Preview handoff prompt (default): node dist/cli.js handoff
• Send relay via OpenClaw: node dist/cli.js handoff --send
• Skip auto-branch creation: node dist/cli.js handoff --no-branch

Repos Under Discussion

| Repo | Role | URL | |------|------|-----| | OpenClaw | Agent runtime (gateway, sessions, channels) | https://github.com/openclaw/openclaw | | Mission Control (Bhanu) | Dashboard variant — FastAPI + Postgres | https://github.com/abhi1693/openclaw-mission-control | | Mission Control (Manish) | Dashboard variant — Convex, hook-based | https://github.com/manish-raana/openclaw-mission-control | | OpenClaw Studio | Dashboard variant — gateway-first, no DB | https://github.com/grp06/openclaw-studio | | OneContext | Unified agent context layer | https://github.com/TheAgentContextLab/OneContext |