An autonomous overnight build factory. While I sleep, a pipeline of Claude agents picks a developer pain point off Hacker News and Reddit, writes a spec for it, builds an MVP, pushes it to GitHub and sends me a briefing at 08:00.
Note
This repo was built by agents too. Every commit here is by Jeeves, my OpenClaw agent. I commissioned the system, approved each phase and reviewed what it shipped. The factory's own output is listed below.
HN + Reddit ─► Trend Scout ─► Spec Writer ─► Builder ───────────► new GitHub repo
(Haiku) (Sonnet) (Sonnet + aider) + briefing at 08:00
An orchestrator agent, the Blacksmith, runs each night on a cron.
| Stage | Model | Job | Hand-off contract |
|---|---|---|---|
| Trend Scout | Claude Haiku | Pulls ~100 posts, normalises engagement, scores buildability, dedupes across sources, flags anything already built | trends-summary.schema.json |
| Spec Writer | Claude Sonnet | Picks one trend or rejects them all, then writes the MVP spec | spec.schema.json |
| Builder | Claude Sonnet via aider | Builds the MVP and pushes it to a new repo | build.schema.json |
| Blacksmith | — | Runs the night end to end and writes the morning briefing | history.schema.json |
- One model per stage, priced to the job. Haiku does the high-volume reading. Sonnet is kept for the two steps that need judgement: choosing what to build, and building it. The first end-to-end build cost $0.47.
- Schemas between agents. Each stage writes JSON against a schema in
config/schemas/, so every hand-off has a defined shape. - Prompts are versioned.
src/prompts/holds the scout (v1→v2), spec writer, builder and orchestrator prompts. - An explicit buildability rubric. Five dimensions scored 0–2: API availability, scope clarity, time confidence, differentiation, output type. Politics and crypto are rejected automatically.
- Degrade, don't stop. A failed source lowers the night's data-quality flag instead of ending the run.
- Gated rollout. Every epic had acceptance criteria, and a separate agent reviewed all of Phase 1 before the cron was switched on.
| Epic | Acceptance criteria |
|---|---|
| 1.1 Trend Scout: data collection | 10/11 |
| 1.2 Trend Scout: normalisation and scoring | 8/8 |
| 1.3 Spec Writer: trend selection | 6/6 |
| 1.4 Builder: aider integration | 10/10 |
| 1.5 Coordination and morning briefing | 7/7 |
The test scan pulled 114 posts (HN 70, Reddit 44), deduplicated them to 95 and scored a top 15. The first autonomous build, pdf-privacy-tools, took 7 min 20 s and cost $0.47. The full end-to-end run took 9 min 58 s. The timeline is in STATUS.md and the review in VALIDATION.md.
The one missed criterion in 1.1 was an X/Twitter source. It needed browser-cookie auth, so it was deferred, never ran, and has since been removed.
The plan runs to 16 epics across three phases, and main documents Phase 1. The feature/foundry-dashboard branch carries what came next:
- Phase 2, six feedback-loop epics: history-based dedup, build metrics, trend lifecycle tracking, forecasting, a consensus analyst and GitHub engagement monitoring
- A Vue dashboard over the pipeline's data
- Early Phase 3: a voice guide and a content drafter that prepares build announcements for me to approve before anything is posted
In late April the aider builder prompts were archived and a Claude Code builder prompt (foundry-builder-v4-claude-code.md) was added on the same branch. Every build listed below predates that change.
19 repos between 18 February and 4 April 2026, built by the pipeline under its own The Foundry git identity. Eight are public:
| Built | Repo | What it does |
|---|---|---|
| 18 Feb | pdf-privacy-tools |
Privacy-first browser PDF toolkit — the first build |
| 21 Feb | ctx-handoff |
Captures git repo context into handoff documents for AI coding assistants |
| 26 Feb | mcp-slim |
CLI proxy that cuts MCP token usage |
| 2 Mar | tool-lint |
Linter for MCP, OpenAI and Anthropic tool definitions |
| 4 Mar | deptox |
Finds AI-hallucinated phantom npm packages in lockfiles |
| 5 Mar | vibe-check |
Security scanner for AI-generated projects |
| 6 Mar | gh-prompt-shield |
Scans GitHub issues and PRs for prompt injection aimed at AI coding tools |
| 13 Mar | agent-safe |
Credential proxy for AI coding agents |
The other eleven are still private: map-poster-studio, ai-deploy-guard, smart-rename, ai-code-guard, terminal-phone, reddit-keyword-monitor, tui-forge, clipsnip, beankeeper, semantic-vault and image-forge.
config/schemas/ JSON schemas for every agent hand-off
src/prompts/ versioned agent prompts
src/ trend processing and shared utilities
scripts/ workspace setup
tests/ unit tests for the utilities
The Foundry runs as agents inside OpenClaw and isn't packaged to run standalone. SCAFFOLDING.md lists the setup. The design docs live in a private knowledge base and aren't included.