Problem
Coding agents are fast. Running several products alone still requires an overview: what is approved, at which revision, within which limits – and who decided last?
Solution
App-Leitstand isn't another agent platform. It's the governance layer a solo operator needs to let agents work safely.
- Requirements → reviewed versions with checksum → starter package for the coding agent → stages as tasks → results → approvals.
- Approvals bind to project, action, target, source revision and content; any change invalidates a decision.
- Project factory: generated starter packages (AGENTS.md, stage plan, licence, stack scaffolds) give the agent clear boundaries.
Architecture and decisions
- MCP server (stdio and Streamable HTTP): Claude Code and Claude Desktop read tasks and approvals directly and report results back – one tool source for both transports.
- LangGraph pilot in Python with interrupts, SQLite checkpoints and a deterministic offline test model – labelled honestly, no provider call without a budget decision.
- Online only behind Cloudflare Access plus its own JWT check: two locks.
- Zero-cost architecture: self-hosted tools (Langfuse for tracing) instead of SaaS, Docker Compose instead of installers, ops scripts with dry run as the default.
Status
App-Leitstand already steers my own development, including the neuefische capstone project. An open-source release is in preparation.