An AI workflow system Peter Marsala built and runs every day
MarsalaCC is a self-hosted AI knowledge and workflow system that Peter Marsala designed, built with AI coding assistants, and uses for seamless planning and facilitation. Its AI agents handle routine tasks and check their own sources and health, and the changes that matter wait for a person to approve them.
Peter's role: he designed the system, directs the AI coding work, and runs it day to day.
See it working
-
System Map
Every publicly listed part of MarsalaCC and how it is used today, from in daily use to retired. The page is generated from the system's own records.
marsala.cc/dashboard
-
Chat with AI Peter
AI Peter's Speakeasy: ask an AI version of Peter about his work. It runs on a local open-source model, with guardrails that keep it on topic and stop it from making up credentials or numbers. It can be wrong, so confirm anything important with the real Peter. Conversations are saved.
talkto.petermarsala.com/chat/welcome
-
Peter's portfolio
Background, experience and how Peter works. He edits and publishes it with a browser editor built as part of MarsalaCC, with a live preview, version history and one-click rollback.
petermarsala.com
Also: reduct.us, Reduct, the company Peter co-founded that builds AI knowledge and retrieval systems.
What is inside
- Memory. Past work and decisions are kept in a database, so each session starts with context.
- Answers with checked sources. It looks things up in its own records and automatically checks that each cited line exists and matches the quote.
- AI agents with approval steps. Scheduled agents do routine work; changes that matter wait for a person to approve them.
- Health checks. A health board and automatic watchers flag parts that stop working.
- Pages that update from records. Some documentation, including the System Map, is generated from the system's records instead of written by hand.
Built with Python, SQLite, Model Context Protocol (MCP) tool servers, Windows Task Scheduler, a Cloudflare Tunnel, local models through Ollama, and supporting services in Docker. Most of the code was written by an AI coding assistant (Claude Code) under Peter's direction.
Contact
Email Peter at [email protected] or find him on LinkedIn.