by owainlewis
Build your own software factory with AI agents
# Add to your Claude Code skills
git clone https://github.com/owainlewis/factoryLast scanned: 8/6/2026
{
"issues": [],
"status": "PASSED",
"scannedAt": "2026-08-06T06:30:47.090Z",
"npmAuditRan": true,
"pipAuditRan": true,
"promptInjectionRan": true
}factory is an open-source ai agents skill for AI coding assistants such as Claude Code, Codex CLI, and ChatGPT, built by owainlewis. Build your own software factory with AI agents. It has 174 GitHub stars.
Yes. factory passed SkillsLLM's automated security scan — a dependency vulnerability audit plus prompt-injection heuristics — with no high-severity issues. You can read the full report in the Security Report section on this page.
Clone the repository with "git clone https://github.com/owainlewis/factory" and add it to your Claude Code skills directory (see the Installation section above).
factory is primarily written in Go. It is open-source under owainlewis on GitHub, so you can review or fork the full source.
Yes. SkillsLLM lists many other AI Agents skills you can browse and compare side by side. Open the AI Agents category from the badge at the top of this page, or use the Related Skills and comparison links further down to weigh factory against similar tools.
No comments yet. Be the first to share your thoughts!
Developer preview
Factory is currently in developer preview and is iterating rapidly. THERE WILL BE COMPATIBILITY-BREAKING CHANGES.
Factory runs repeatable software-engineering Work across Git repositories and compute. Operators save prompts as Routines, run them now, or schedule them across one or more repositories. Its Go control plane coordinates Work and a fleet of Workers that provide Pi, Codex, or Claude Code runtimes.
The browser UI provides:
Factory is local-first. Its browser and operator API accept loopback connections only. An optional, separate TLS-authenticated endpoint accepts Workers on remote VMs.
Factory starts with coding agents on machines you control and is intended to scale to elastic cloud execution without rewriting a Routine or splitting its Work history. Existing Routines remain persistent by default. Once the proposed cloud backend is implemented, a manual run will be able to select a compatible elastic profile. Persistent local and VM Workers remain the rich path for subscription authentication, warm repository caches, and inspectable worktrees. A proposed Cloud Run backend adds disposable, API-backed agent containers for bursty and parallel Work. Read the Cloud Run agent backend design for the planned boundary, security model, and rollout.
Read the product vision, the current implementation architecture, and the active Cloud Run backend design. Planned work follows the project workflow. Superseded proposals remain available through the documentation index.
Requirements:
curljustNode.js is only needed when changing the UI. Normal builds use the committed, embedded UI assets.
Managed GitHub repository checkout and Routines that allow the gh tool depend
on the GitHub CLI. Factory does not include a separate GitHub API client.
Install and authenticate gh on the control-plane
host and each eligible Worker host:
gh --version
gh auth status
just build
mkdir -p ~/.factory
cp examples/worker.toml ~/.factory/worker.toml
Edit ~/.factory/worker.toml to select any installed pi, codex, and
claude-code agents. Workers need no repository list. They probe local gh
access and acquire centrally managed GitHub repositories on demand. Then start
the server and Worker:
just run
Open http://127.0.0.1:7337.
Database migration 27 converts supported pre-launch Definitions, schedules, and execution history into Routines and Work after the database is backed up. Unsupported legacy provider admission is blocked and reported instead of being silently discarded.
One Worker has one stable identity, a configurable set of coding-agent capabilities, and a pool of independent sessions. The pool defaults to ten slots. A single local Worker can advertise Pi, Codex, and Claude Code:
FACTORY_WORKER_CONFIG=~/.factory/worker.toml \
~/.factory/bin/factory-worker
See the local guide for a complete setup and the worker guide for runtime and worktree behavior. Use the remote VM guide to enroll a Worker outside the server host. Tagged binary installation, upgrades, compatibility, rollback, and release verification are covered by the release guide.
Browser
|
| HTTP + JSON
v
Go control plane
SQLite, Routine scheduler, Work admission, embedded UI
^
| registration, claim, heartbeat, events, completion
|
Go Workers
one identity + ready runtime capabilities + N agent slots + repository cache
|
+-- Pi CLI
+-- Codex CLI
`-- Claude Code CLI
The control plane owns durable coordination. Workers own execution and Git worktrees. Workers poll the API, so the system does not require WebSockets or inbound connections to worker hosts.
The future execution model keeps three choices independent:
Execution backend Agent runtime Provider and model
----------------- ------------- ------------------
Persistent Worker + Pi, Codex, Claude Code + subscription or API
Cloud Run Job + Pi, Codex, Claude Code + API-backed model
Cloud Run is a proposed execution backend, not a DeepSeek-specific product. For example, the completed experiment ran the Pi runtime in Cloud Run with DeepSeek V4 Flash through OpenRouter. Cloud execution is managed and elastic, but not infrastructure-free, infinitely scalable, or a hostile-code sandbox.
All default state is below ~/.factory:
~/.factory/
bin/
server/factory.sqlite3
config.toml optional control-plane bootstrap only
worker.toml
workers/
Read the architecture for the contracts and security boundaries.
Implemented:
Designed but not implemented: a unified factory CLI and an elastic Cloud Run
agent backend.
See the documentation index for current guides and proposed designs.
Backend:
just test
just vet
UI:
just ui-install
just ui-check
After changing the UI, rebuild the committed assets:
just ui-build 0
See CONTRIBUTING.md for the full check set.