ai-workspace

ai-workspace Work above your projects, not inside them.

A workspace layer that turns Claude Code from an autocomplete into a team. Agents with enforced permissions run in parallel across repositories. You orchestrate.

The problem

Typical setup: open a repo, run claude, work inside it. Claude sees only that repo.

Real tasks don't work that way. "Add a feature" touches the backend, frontend, and database. "Set up staging" means 12 services and a dozen config files. Claude locked inside one repo doesn't see the system.

The pattern

Claude opened at the workspace root sees all projects in tmp/. Agents with narrow tools lists can't exceed their permissions — not by instruction, but by enforcement.

Two modes

Orchestrator — task spans multiple services:

Main chat coordinates. Agents work in parallel.

Isolated agent — large task in one repo, keep main branch clean:

Quick start

/start detects your project automatically — reads existing CLAUDE.md / AGENTS.md, infers stack from Makefile / go.mod / package.json, configures permissions.

Agent permissions are architecture

An agent's tools frontmatter is enforced by Claude Code — it's not a suggestion.

researcher has no Write or Edit. It physically cannot modify a file. Add your own agents by dropping a .md into .claude/agents/.