Context
opcode (formerly Claudia) is an open-source desktop app and toolkit for Claude Code, licensed under AGPL-3.0. I didn’t build it from scratch: I extended it and rewrote large parts of it for the way I work with coding agents every day.
The problem I was solving is a platform problem, not an AI one. Once several agents work on the same project, you need the same things you need for any shared platform: guardrails, cost limits and governance.
What I added
- Managers and workers. A manager chat plans the work and starts a worker chat for each task. Workers ask their manager for routine go-aheads, so only what really needs me reaches me.
- Caps, budgets and accounts. Each profile has a ceiling for model and effort, so a simple task can’t run on the most expensive settings. Budget stop lines halt work before a usage limit is hit, and work can switch between accounts.
- Locks and an owner inbox. Agents claim files before they edit them, so two agents never change the same file at once. Decisions only I may make go to an inbox as cards instead of being guessed, and risky commands trip a LUKE-ONLY stop that waits for my approval.
- Rules and trackers. Rule files (
MANAGER.md) say what agents may decide on their own. Profiles set the area and limits for each kind of worker, tracker files keep the state of the work, and an AFK mode lets work continue while I’m away.
Why it matters
It’s the same thinking I use on infrastructure: least privilege, budgets and alerts, locks against concurrent changes, and a clear line between what automation may do and what needs a human.
Status
My version isn’t public yet. I plan to publish it, in line with the original project’s licence. The upstream project is at github.com/winfunc/opcode.