Janis started from a simple premise: Claude Code already knows how to work — what it lacks is orchestration. Instead of building a complex platform around the model, Janis is a bash script that connects the pieces that already exist: the Claude Code CLI, the filesystem, CLAUDE.md files as project context, and launchd for scheduling. The result is an autonomous work orchestrator that runs sprints, manages multiple projects, and operates unsupervised overnight.
Features
Philosophy
The central architectural decision in Janis is the deliberate refusal of complexity. There is no database, no server, no graphical interface. State lives in markdown files. Each project's context is a CLAUDE.md at the repository root. Communication between the operator and the system happens through the filesystem — an inbox/ folder with daily briefings, notes, and logs. This simplicity is not a limitation: it is the recognition that the language model is already sophisticated enough, and what it needs is clear structure, not heavy infrastructure.
How it works
Janis operates in three modes. Interactive mode is a normal Claude Code session with loaded context — projects, priorities, history. Autopilot mode executes structured sprints: the model plans milestones, executes each one, validates the result, and moves to the next, all without human intervention. Nightshift is a nightly session scheduled via launchd that processes background tasks — market research, drafts, competitive analysis — while the operator sleeps. All three modes share the same context through CLAUDE.md and a feedback system that records operator approvals and corrections.
What's next
Janis is available as open source and fully functional for anyone who wants autonomous work orchestration without complexity. Meanwhile, a successor is in development — a more autonomous agent architecture featuring persistent memory via a temporal knowledge graph, a learned value function that improves decisions over time, and continuous consciousness across sessions. Janis proved that 800 lines of bash are enough to orchestrate serious work. The next generation asks what becomes possible when the agent can also remember, learn, and act on its own initiative.