Start an agent team in 30 seconds. Learn from every run.

Drop-in Agent Teams enhancement for Claude Code — team templates, intent bundles, smart hooks, and custom agents. Zero config required.

Neo does one thing well: take Claude Code's raw Agent Teams API and make it usable without a PhD in orchestration. Install it, prompt naturally, get a team.

# 1. Clone
git clone https://github.com/Hybirdss/claude-agent-neo.git ~/dev/claude-agent-neo

# 2. Install
cd ~/dev/claude-agent-neo && ./install.sh

# 3. Use — just describe your task
"Build a login page with frontend, backend, and security review"

The auto-team-spawner hook detects complexity and recommends the right template. The model executes — you stay in control.

git clone https://github.com/Hybirdss/claude-agent-neo.git ~/dev/claude-agent-neo
cd ~/dev/claude-agent-neo && ./install.sh

The installer symlinks files into ~/.claude/ — existing config is not overwritten.

# Install only specific modules
./install.sh --modules agents,hooks,templates,bundles

# Verify the install
bash scripts/health-check.sh

One JSON file = one complete team with tasks, file scopes, dependencies, and spawn order.

{
  "name": "fullstack-build",
  "description": "Frontend + Backend + Reviewer team",
  "stagger_seconds": 5,
  "members": [
    { "name": "frontend", "agent": "parallel-worker", "fileScope": ["src/components/**"], "worktree": true },
    { "name": "backend",  "agent": "parallel-worker", "fileScope": ["src/api/**"],        "worktree": true },
    { "name": "reviewer", "agent": "code-reviewer",   "planModeRequired": true }
  ],
  "tasks": [
    { "subject": "Frontend implementation", "owner": "frontend" },
    { "subject": "Backend implementation",  "owner": "backend" },
    { "subject": "Integration review",      "owner": "reviewer", "blockedBy": ["1", "2"] }
  ]
}

stagger_seconds — delay between spawning each agent. Default: 5s. Prevents context collisions when agents start simultaneously.

Drop a JSON file in team-templates/ to add your own.

Saves session state to state.json on stop. Next session restores: current phase, active files, last commit, blockers, next actions.

Injects leader context into each teammate's initial prompt — project phase, active files, current goal, recent commits. Teammates start informed.

Logs every tool call to ~/.neo/logs/session-{date}.jsonl:

{"ts": "2026-04-09T10:23:01Z", "agent_id": "leader", "tool": "Edit", "input_summary": "{\"file_path\": \"src/auth.ts\"..."}

Use logs to debug failed runs, measure cost per agent, or replay what happened. NEO_AGENT_ID env var tags which agent made the call.

When a team agent session ends, injects a reminder: STATUS: DONE/BLOCKED/NEEDS_CONTEXT. Only active when NEO_AGENT_ID is set — silent for the leader.

Detects complex tasks (file counts, fullstack keywords, migration scope) and recommends the best template — including stagger_seconds for safe spawn order. The model decides whether to follow it.

Records every team event (task changes, messages, file writes) into a timeline JSON. Use for debugging failed runs and comparing template efficiency.

Extracts lessons from finalized replays and stores them per-template. Next run of the same template gets the lessons injected — file conflicts, slow tasks, review overload patterns.

Monitors CI after git push. Re-wakes the model with failure details if CI fails. Exit code 2 triggers re-wake.

Classifies prompt intent and injects relevant skills/agents. Example: "review this code" → review-fix bundle loads with review, investigate, and critic skills.

All tool calls are logged to ~/.neo/logs/session-{date}.jsonl by the session-logger hook.

# Tail live activity
tail -f ~/.neo/logs/session-$(date +%Y-%m-%d).jsonl | jq .

# Count tools used per agent today
jq -r .agent_id ~/.neo/logs/session-$(date +%Y-%m-%d).jsonl | sort | uniq -c

Each line: ts (ISO8601), agent_id (NEO_AGENT_ID or "leader"), tool, input_summary (first 100 chars).

Read-only, no CLAUDE.md loading, fast startup. For lookups and web searches.

Reviews without modifying. Reports findings with severity and file location.

Team worker with file scope boundaries. Reports progress via TaskUpdate.

Verifies install: symlinks exist, hooks registered in settings.json.

bash scripts/health-check.sh
# PASS agents/ symlink
# PASS bundles/ symlink
# PASS hooks registered

Live terminal dashboard: team status, task progress, member activity.

watch -n 1 bash scripts/team-dashboard.sh [team-name]

claude-agent-neo/
├── agents/                    # Custom agent definitions
├── bundles/                   # Intent bundle manifests
├── team-templates/            # Team composition templates (with stagger_seconds)
├── hooks/
│   ├── project-state.py       # Stop: save state.json
│   ├── context-bridge.py      # SubagentStart: inject leader context
│   ├── session-logger.py      # PostToolUse: JSONL tool log → ~/.neo/logs/
│   ├── completion-guard.py    # Stop: STATUS reminder for team agents
│   ├── async-watcher.sh       # PostToolUse: CI monitoring
│   ├── bundle-loader.py       # UserPromptSubmit: intent classification
│   ├── auto-team-spawner.py   # UserPromptSubmit: auto team recommendation
│   ├── team-replay.py         # PostToolUse: team event recording
│   └── team-memory.py         # Stop: cross-session team learning
└── scripts/
    ├── install.sh             # Symlink installer (--modules support)
    ├── health-check.sh        # Post-install verification
    └── team-dashboard.sh      # Live team status dashboard

• Claude Code CLI with Agent Teams support
• CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 in settings.json env

MIT