A local control plane for AI-assisted software development. No magic-from-a-socket — instead a structured production line of wizard, supervised runs, Git discipline and an automatic quality gate. Reproducible, traceable, sign-off-able.
New · v0.30.0 Some gaps you don't close — you take them apart. „Human working time“ stays unmeasurable: the factory sees how long something waited, not whether anyone was working on it, and an open browser tab is not work. What could be separated out is now reported: the interventions (how often a human had to decide) and the number of deciders. The effort figure beside them is asked for, not measured — voluntary, and always accompanied by how many decisions it actually covers. Before that, v0.29.0 brought rollback detection. All news →
The six-step wizard only asks what's needed for a concrete
project: target platform, backend, frontend(s), stack-specific
questions, quality-gate toggles, summary. 18 templates —
six backends from Spring Boot to FastAPI, eleven frontends from
React to JavaFX, plus importing an existing repository. The
result: PROJECT.md, AGENTS.md and an
initial prompt you can preview before starting.
Run orchestration with live logs, token stream and
container sandbox. Every agent runs with hard limits
(--cpus 2 --memory 4g --network=none), writes
only to the workspace, every change is a commit. Browser
notification when you need to make a decision.
Inline diff before every approval. Automatic quality gate from six reviewer roles — Security, Architecture, Hallucination, plus two vendor reviewers. No "Approve / Reject" blind flight, but findings sorted by confidence.
Reference architecture, agent orchestration, guardrails and SDLC integration — as an architecture document, not a tool comparison. If you want to know why the factory is cut this way, start here.
Coding agents are already good at solving individual problems. They are bad at sticking to structure, cutting commits sensibly, getting reviews and checking their own output. The Software Factory takes those steps over — not the creative part.
The result: AI-assisted development that an architect can hand to a client without cleaning up the trail afterwards.
The Architect gets the strongest available frontier model, the Reviewer a lean, fast one — saving tokens where quality isn't hurt and giving the strong model to the strategic roles. The concrete mapping comes from the model registry and follows new releases automatically.
Before every run the platform writes
.claude/settings.local.json and
.claude/agents/<role>.md into
the workspace. Subagents are natively delegable
without touching the CLI.
Optional: every agent runs in an ephemeral Docker
container with hard limits,
--read-only root and
--network=none. The answer to "is it
running on my filesystem?" is: no, only on its
mount.
At WAITING_FOR_APPROVAL the platform
shows the final git diff. You see what
the agent changed before you press the button.
Wizard step 4 shows a token and EUR estimate — before the first run, without a vendor call. JTokkit computes locally.
Software architect with over three decades between mainframe (z/OS, Cobol, REXX) and cloud (Java/Spring, AWS, Azure, Kubernetes, microservices). In the frontend with Apache Wicket and Angular, in mobile founder of a German-speaking Xamarin community.
Certifications (selection): Claude Partner Network (Anthropic), CKAD — Certified Kubernetes Application Developer (CNCF), CCA-Foundations, AWS Certified Cloud Practitioner, Microsoft Certified: Azure Fundamentals (AZ-900), iSAQB CPSA-F (software architecture) and Oracle Certified Professional 12c.
The Agentic Software Factory is my attempt to embed AI agents into the software lifecycle in a way that makes them productive even under mid-market and compliance constraints — structured, traceable, sign-off-able rather than blindly generated.
