Claude Litter is an enhancement suite for Claude Code agent teams. Its core is the swarm-loop plugin — an orchestrated multi-agent loop that decomposes complex goals into parallel subtasks, drives autonomous iteration with hook-based safety, and exits only when a user-defined completion promise is fulfilled. Alongside the plugin, a terminal UI gives you a live control plane for managing teams, agents, tasks, and swarm executions without leaving the terminal.
Why Claude Litter?
Claude Code's native agent teams are powerful, but undercooked. The swarm-loop plugin adds structured orchestration on top of teams — iterative loops with safety classifiers, stuck detection, and progress tracking — while the TUI provides a simple interface that can assistant with managing the.
| Pain | Claude Litter |
|---|
| Orchestration | Manual agent coordination | Swarm loop with autonomous iteration |
| Team overview | cat ~/.claude/teams/*/config.json | Live sidebar with status indicators |
| Agent transcripts | Buggy interface | Tabbed sessions with full history |
| Task tracking | Manual JSON edits | Filterable panel with inline editing |
| Messaging | Write to inbox JSON files | Compose form with broadcast support |
| Agent lifecycle | Shell commands, manual PID tracking | Spawn, kill, detach, reattach in-UI |
| Swarm monitoring | cat .claude/swarm-loop/*/state.json | Live progress, phase, and log streaming |
Features
Swarm Loop Plugin
The core of Claude Litter — orchestrated multi-agent iteration for Claude Code.
- Iterative orchestration with profile-driven cycles (default 7-step, leanswarm 4-concern, deepplan 5-phase planning, async background agents)
- Hook-based safety — bash classifier, teammate idle gate, task completion/creation gates, sentinel timeout recovery
- Completion promise system with optional shell verification (
--verify)
- Iteration control — soft budget checkpoints, hard min/max floors and ceilings, stuck detection with permission-aware escalation
- Configurable via
.claude/swarm-loop.local.md — classifier model/effort, teammate isolation (shared or worktree), max count, notifications
See Swarm Loop Plugin for full documentation.
Terminal UI
A control plane for managing Claude Code agent teams and swarm executions.
- Team sidebar with live status indicators (active / partial / inactive) and colored agent badges
- Tabbed sessions with full conversation transcript history loaded from JSONL files
- Swarm panel with live iteration status, phase, autonomy health, progress bars, and incremental log streaming
- Task panel with filtering (pending / in-progress / completed / blocked), sorting (ID / status / owner), and inline editing
- Message panel with inbox view, broadcast view, and compose form
- Agent lifecycle — spawn with model selection, configure, duplicate cross-team, kill, detach, reattach
- Kitty terminal pop-out: open agents in splits, tabs, or new windows
- Live filesystem watching — changes to team/task/inbox/swarm JSON files refresh the UI automatically
- 10 built-in themes, vim mode, command mode with Tab autocomplete, debug logging
Swarm Loop Plugin
The swarm-loop plugin is an orchestrated multi-agent iterative development system for Claude Code. It decomposes complex goals into parallel subtasks, creates a persistent agent team, and drives autonomous iteration until a user-defined completion promise is fulfilled.
How It Works
