Get from zero to orchestrating AI agents in 5 minutes
Andrej Karpathy said it best: you could be 10X more powerful with AI agents if you could just string together what's available. JAT is that string.
# Linux (Arch/Manjaro)
sudo pacman -S tmux sqlite jq nodejs npm
# Linux (Debian/Ubuntu)
sudo apt install tmux sqlite3 jq nodejs npm
# macOS
brew install tmux sqlite jq nodeWhy these?
tmux- Agent sessions run here (IDE tracks them)sqlite3- Agent Registry database (coordination)jq- JSON processing (tool outputs)node/npm- IDE and browser automation
# One-line installer (works on Linux and macOS)
curl -fsSL https://raw.githubusercontent.com/joewinke/jat/master/install.sh | bashThis symlinks all tools to ~/.local/bin/:
jt-*- JAT Tasks CLIam-*- Agent Registry (4 commands)browser-*- Browser automation (11 commands)db-*- Database tools (4 commands)jat-*- JAT signals and launchers
Reload your shell:
source ~/.bashrc # or ~/.zshrc on macOSVerify:
am-whoami # Should show "Not registered" (expected)
jt --version # Should show version numbercd ~/code/myproject
jt initAnswer the prompts (or press Enter for defaults):
- Project name:
myproject - Default priority:
2(Medium) - Task prefix:
mp(e.g., mp-001, mp-002)
This creates .jat/ directory - git-backed task tracking.
jatThis:
- Checks dependencies (npm install if needed)
- Starts SvelteKit dev server (port 3333 by default)
- Opens browser to http://localhost:3333
You should see:
- Empty Work page (no active sessions yet)
- Navigation: Tasks, Work, Files, Servers, Agents, Config
- Go to Tasks page
- Click "Create Task" button
- Fill in:
- Title: "Add authentication"
- Type:
task - Priority:
P2(Medium)
- Click "Create"
- Go to Work page
- Click "Start Next" dropdown → Select your new task
- Watch agent spawn in tmux session
# Single agent
jat-myproject
# Multiple agents in auto-attack mode
jat myproject 4 --autoWhat happens:
- Tmux session created:
jat-{AgentName} - Claude Code starts in session
- Agent runs
/jat:start(picks highest priority task) - IDE shows live terminal output
- Session state updates in real-time
What you have:
- ✅ Visual IDE showing all agent sessions
- ✅ Task management (create, assign, track)
- ✅ Live terminal output per agent
- ✅ Code editor (
/filespage) - ✅ Server controls (
/serverspage) - ✅ Token usage tracking
Go to Config → Review Rules:
The Review Rules Matrix:
P0 P1 P2 P3 P4
bug review review review auto auto
feature review review auto auto auto
task review auto auto auto auto
chore auto auto auto auto auto
What this means:
- P3-P4 bugs: Agent auto-completes → spawns next task
- P0-P2 bugs: Agent asks for review before completing
- Click cells to toggle between auto/review
Why configure this?
- Lets agents work overnight on low-priority items
- You review high-priority completions
- Balances autonomy with control
Go to Config → Automation → Presets:
Install these presets:
- ✅ API Overloaded Recovery (auto-retry with backoff)
- ✅ Rate Limit Recovery (wait 60s, retry)
- ✅ Network Error Recovery (detect ECONNREFUSED, retry)
- ✅ Auto-Continue Prompts (send Enter on "Continue?" prompts)
Why these help:
- Agents recover from transient errors automatically
- You don't babysit stuck sessions
- 3am rate limit? Handled.
- Tasks page → Create Task
- Type:
epic - Title: "Add user authentication system"
- Description:
- [ ] Design auth flow - [ ] Add login endpoint - [ ] Add registration endpoint - [ ] Add password reset - [ ] Write tests - [ ] Update documentation - Click Create
- Click the Epic's title → Opens detail drawer
- Click "Launch Epic Swarm" button
- Configure:
- Execution mode: Parallel
- Max concurrent: 4
- Review threshold: P0-P1 (only review critical items)
- Auto-spawn blocked: Yes (wait for dependencies)
- Click "Launch"
- IDE converts each
[ ]item into a subtask - 4 agents spawn simultaneously
- Each picks a subtask (ordered by priority)
- As tasks complete, agents auto-proceed to next task
- Progress bar shows completion
When to intervene:
- Agent shows "NEEDS INPUT" badge → click to answer
- Agent shows "REVIEW" badge → check work before completing
- Error in terminal → automation rules try recovery first
Cause: Not running in tmux
Fix:
# Don't run:
claude "/jat:start" # ❌ No tmux, IDE can't track
# Do run:
jat-myproject # ✅ Launcher creates tmux sessionCause: Haven't run jt init in your project
Fix:
cd ~/code/myproject
jt init
# Refresh IDECause: npm dependencies not installed
Fix:
cd ~/code/jat/tools/browser
npm installCheck usage:
- IDE → Agent card → Shows tokens + cost today/week
- Adjust review rules to require more human review
- Set max concurrent lower (fewer agents = less cost)
Typical costs:
- 100K tokens/day ≈ $0.30/day
- 1M tokens/day ≈ $3/day
- 10M tokens/day ≈ $30/day
- Launch one agent on one task
- Watch it work in IDE
- Review and complete
- Configure review rules matrix
- Let agents auto-complete low-priority tasks
- Wake up to finished backlog
- Create epic with 10+ subtasks
- Launch 4-8 agents in parallel
- Track progress in real-time
- Write custom automation rules (regex patterns)
- Template variables with capture groups
- Project-specific error handling
- Initialize multiple projects (
~/code/project1,~/code/project2) - IDE shows all projects
- Swarm attack across projects
Every route with a list, table, or grid supports vim-style j/k navigation. Press ? anywhere for the full route-specific overlay.
| Shortcut | Action | Context |
|---|---|---|
j / k (or ↓ / ↑) |
Navigate list / rows / cards | Any route |
Enter |
Open focused item | Any route |
? |
Show shortcut overlay | Any route |
Esc |
Clear focus / close overlay | Any route |
Alt+N |
Create new task | Global |
Alt+E |
Open Epic Swarm modal | Global |
Alt+S |
Start next task | Global |
Alt+A |
Attach to session | Hovered session (Work page) |
Alt+K |
Kill session | Hovered session |
Alt+1 to Alt+9 |
Jump to session | Work page |
Cmd+K |
Command palette | Global |
Full per-route reference: ide/docs/keyboard-navigation.md.
jat # Launch IDEjat init # Auto-discover ~/code/* projects
jat add <path> # Add project manually
jat list # Show all projectsjt ready # Tasks ready to work
jt create "Title" --priority 1 # Create task (P1 = High)
jt show <id> # View task details
jt close <id> # Close completed task
jt dep add <task> <blocker> # Add dependencyam-whoami # Check your agent identity
am-agents # List all registered agents/jat:start Pick task, reserve files, begin work
/jat:complete Verify, commit, close task, end session
/jat:tasktree Convert PRD to structured tasks
/jat:verify Run tests, lint, security checks
/jat:doctor Diagnose and repair JAT setup
- README.md - Overview and philosophy
- GETTING_STARTED.md - Complete tutorial
- CLAUDE.md - Complete developer reference
- shared/signals.md - Signal system
- shared/automation.md - Automation rules
- shared/tools.md - All 40+ tools
- Check GitHub Issues
- Read troubleshooting in CLAUDE.md
- Join community discussions
Karpathy said programmers could be 10X more powerful if they could just string together what's available.
You just strung it together.
- ✅ Agents orchestrated
- ✅ IDE running
- ✅ Tasks tracked
- ✅ Automation configured
You're ready to claim the 10X boost. Happy orchestrating! 🚀