This is how we operate.
One loop runs the company: it picks the opportunity, builds the product, drafts the marketing, and closes its own books. A human holds the gates. Everything below is the real system, not a demo.
Eleven steps, two human gates.
- 01CEO decisionmoltfire
Opens the cycle and authorises the research run.
- 02Researchps-researcher
Pulls real complaint and trend signals, ranks opportunities.
- 03Brainstormps-brainstormer
Turns the top opportunity into distinct, concrete product concepts.
- 04CEO pickmoltfirehuman gate
Scores the concepts and picks one. In gated mode the cycle parks here until a human approves the pick in DashClaw.
- 05Build briefps-builder
Writes an executable build brief for the chosen product.
- 06Buildmoltfire
Dispatches the managed builder, a headless Claude Code agent, and waits on the result.
- 07Marketingps-marketer
Drafts landing copy and launch posts. Drafts only, never publishes.
- 08QAps-qa
Scores the drafts for honesty, fluff, and broken claims. Can block the cycle.
- 09Outreachps-salespersondrafts only
Queues personalised outreach as pending approval. The loop never auto-sends; a human approves every send.
- 10Financeps-finance
Produces the cycle P&L from real LLM cost data.
- 11CEO closemoltfire
Writes the outcome summary, decides the reinvest direction, closes the books.
Dry-run cycles exercise the whole loop with outreach disarmed. Live cycles queue outreach as drafts pending approval. Neither mode can send anything on its own.
Mission Control, as it looks today.
The operator's view of the company. These are current screenshots of the real cockpit, not mockups. Click to zoom.
Four lanes, one company.
- Company loopMoltFire (CEO), Researcher, Brainstormer, Builder, Marketer, QA, Salesperson, Finance
Runs the 11-step cycle: pick, build, market, close the books.
- Sales laneProspector, Hygiene, Researcher, Outreach, Orchestrator
Finds and scores prospects, researches constraints, drafts sequences for human approval.
- Content laneResearch, Outline, Draft, Publish
Produces the blog end to end; a human revises and approves before anything deploys.
- Build laneManaged builder (headless Claude Code)
Ships the product the cycle picked, gated and logged with its exit code.
Autonomous, not unaccountable.
DashClaw gates
Risky actions are intercepted before they run and held for approval. A parked cycle resumes only when a human approves the action in DashClaw.
Drafts only, never auto-send
Outreach, marketing posts, and support replies are queued as drafts. Every external send requires an explicit human approval.
Kill switch
One flag halts every company run. Cycles check it between steps and stop cleanly.
Full audit trail
Every agent action is recorded with intent and outcome, in the governance ledger and the cycle artifacts.
What it stands on.
One shared database
Agents read and write the same Postgres. Cycle artifacts are file-canonical with a database mirror, so the audit trail survives redeploys.
Neon Postgres + cycle artifacts
Governed at the seam
Agents call DashClaw before acting and record after. Approval gates, risk scoring, and the kill switch live outside the agents themselves.
DashClaw
Managed build agent
The builder runs as a headless Claude Code agent on a fixed subscription seat. The loop dispatches it; governance gates it; every build logs its exit code.
Claude Code, gated dispatch
Cost-disciplined LLM layer
Bulk work runs through the batch API, prompts are cached, each task routes to the smallest adequate model, and true per-call cost lands in the ledger.
Batch API, caching, model routing
The sales lane still runs its own state machine for prospects; the company loop is the spine everything else hangs off.