Stop Coding Alone! ClawTeam's Agent Swarm Does the Work

Stop Coding Alone! ClawTeam's Agent Swarm Does the Work

What if your AI assistant could hire its own team?

Picture this: It's 2 AM. You're staring at a sprawling codebase, knowing that building this full-stack application requires auth, database design, API architecture, frontend components, and comprehensive tests. You've got Claude Code running, but it's drowning in context, switching between files, losing track of dependencies. Every prompt feels like herding cats. Meanwhile, your competitor just shipped a feature that took them three days—because they weren't coding alone.

The dirty secret of 2025? Solo AI agents are already obsolete.

Top engineering teams aren't wrangling single chatbots anymore. They're deploying agent swarms—coordinated teams of specialized AI workers that divide complex projects, communicate autonomously, and converge on solutions faster than any human could orchestrate. The infrastructure gap between elite developers and everyone else is widening by the week. And until now, building that infrastructure meant wrestling with Docker^{↗ Bright Coding Blog} containers, Redis queues, message brokers, and orchestration YAML that nobody wants to write.

Enter ClawTeam—the open-source agent swarm intelligence framework that transforms one command into full automation. No cloud infrastructure. No complex configs. Just pip install clawteam, tell your leader agent the goal, and watch an entire team of specialized AI workers self-assemble, delegate tasks, and ship results while you sleep.

This isn't science fiction. This is HKUDS/ClawTeam, and it's about to change how you build software forever.

What is ClawTeam?

ClawTeam is an agent swarm intelligence framework developed by HKUDS (Hong Kong University Data Science Lab) that enables AI coding agents to self-organize into collaborative teams. Launched publicly in March 2026, it represents a fundamental evolution in AI-assisted development: from solo agents executing isolated tasks to swarms of agents thinking and working together with collective intelligence.

The project's tagline says everything: "One Command → Full Automation."

Here's the breakthrough: ClawTeam doesn't just run multiple agents side by side. It creates a complete organizational structure where a leader agent spawns specialized sub-agents, each with dedicated environments, assigns tasks with intelligent dependency management, monitors progress in real-time, and dynamically reallocates resources based on performance. The agents communicate through an inbox system, report status automatically, and even cross-pollinate insights between parallel experiments.

What makes ClawTeam genuinely disruptive is its universal compatibility. Unlike framework-locked alternatives that force you into specific agent ecosystems, ClawTeam works with any CLI agent: Claude Code, OpenAI Codex, OpenClaw, nanobot, Cursor, Kimi CLI, or your custom scripts. The only requirement? The agent can execute shell commands. That's it. No SDK integration, no API wrappers, no vendor lock-in.

The project has gained rapid traction because it solves the coordination bottleneck that plagues every serious AI-assisted project. When Karpathy demonstrated autonomous ML research with multiple experiments, he exposed what's possible—but also how painful manual orchestration becomes. ClawTeam automates that orchestration layer entirely, collapsing months of infrastructure setup into a single pip install.

Key Features That Make ClawTeam Insane

🦞 Agent Self-Organization with Zero Manual Setup

The killer feature? When you spawn a worker agent, ClawTeam auto-injects a coordination prompt directly into its context. Workers automatically know how to check their task list, update status, message the leader, and report idle state. No training. No custom prompts. No framework-specific code. Any CLI agent becomes a team player instantly.

🌳 Git Worktree Isolation for Conflict-Free Parallelism

Every agent receives its own git worktree on a separate branch (clawteam/{team}/{agent}). This isn't container-based isolation—it's better. Real branches mean real diffs, clean merge workflows, and native git operations. You can checkpoint agent progress, merge successful experiments back to main, and clean up worktrees without Docker overhead. Parallel agents won't step on each other's code, and you retain full version control transparency.

📋 Intelligent Task Dependencies with Auto-Unblocking

Tasks aren't just lists—they're dependency graphs. Create a task with --blocked-by T1,T2 and it stays frozen until predecessors complete. When an agent finishes work, dependent tasks auto-unblock and notify waiting workers. The leader can block on entire pipelines with clawteam task wait --timeout 300, enabling sophisticated CI-like workflows without any CI configuration.

💬 Inter-Agent Messaging: File-Based or ZeroMQ P2P

Agents communicate through point-to-point inboxes with send, receive, peek, and broadcast operations. The default file-based transport requires zero infrastructure—just a shared filesystem. For low-latency scenarios, switch to ZeroMQ P2P with automatic offline fallback. No Redis. No message queues. No database. All state lives in ~/.clawteam/ as atomic JSON files.

📊 Multi-Modal Monitoring: Terminal, Tiled tmux, or Web UI

Watch your swarm through clawteam board show for terminal kanban, clawteam board attach for a tiled tmux view showing all agents working simultaneously, or clawteam board serve --port 8080 for a real-time Web UI. The tiled view is genuinely mesmerizing—eight agents across eight GPU terminals, each coding independently, visible at a glance.

🎪 TOML Team Templates for Instant Domain Teams

Pre-built templates define complete team archetypes. The built-in AI Hedge Fund template spawns a 7-agent investment committee with one command: portfolio manager, five specialized analysts (value, growth, technical, fundamentals, sentiment), and a risk manager. Create templates for DevOps^{↗ Bright Coding Blog}, content production, scientific research—any domain with repeatable team structures.

Use Cases Where ClawTeam Absolutely Dominates

🔬 Autonomous ML Research at Scale

Based on Karpathy's autoresearch framework, ClawTeam enables 8 agents × 8 H100 GPUs running thousands of experiments with zero human intervention. The leader reads a program.md, spawns specialized researchers each exploring different hyperparameter dimensions (model depth, width, learning rates, batch sizes), checks results every 30 minutes, identifies breakthroughs, and cross-pollinates findings to fresh agents.

Real result: 2,430+ experiments, val_bpb improved from 1.044 to 0.977 (6.4% gain), approximately 30 GPU-hours. Months of manual tuning compressed into hours of intelligent automation.

🏗️ Full-Stack Application Development

Tell Claude Code: "Build me a full-stack todo app with auth, database, and React^{↗ Bright Coding Blog} frontend." The leader self-organizes a 5-agent team with dependency chains: architect designs API schema → backend1 implements JWT (blocked by T1) → backend2 builds PostgreSQL^{↗ Bright Coding Blog} models (blocked by T1) → frontend builds React UI → tester writes integration tests (blocked by T2, T3, T4). Each agent works in isolated git worktrees, messages deliverables through inboxes, and the leader merges everything to main.

💰 AI Hedge Fund Operations

One command launches complete investment analysis: clawteam launch hedge-fund --team fund1 --goal "Analyze AAPL, MSFT, NVDA for Q2 2026". The portfolio manager spawns five analyst agents with distinct strategies, collects their signals, routes to a risk manager for position limits, and synthesizes final buy/sell/hold decisions. All inter-agent communication happens through structured inbox messages with automatic consolidation.

🎪 Custom Domain Teams

Scientific research labs spawn literature review, hypothesis generation, experimental design, statistical analysis, and paper-writing agents. Content studios deploy research, drafting, editing, fact-checking, and SEO optimization agents. Business operations teams create sales, marketing, support, and analytics agent committees. The TOML template system makes any repeatable team structure instantly reproducible.

Step-by-Step Installation & Setup Guide

Prerequisites

Before installing, verify your environment meets these requirements:

• Python^{↗ Bright Coding Blog} 3.10+
• tmux installed and functional (tmux -V)
• A CLI coding agent already working standalone (Claude Code, Codex, nanobot, etc.)

Critical: If your agent CLI doesn't run correctly by itself, clawteam spawn cannot fix it. Test first:

tmux -V
claude --version  # or: codex --version, nanobot --help

Installation Methods

Option 1: PyPI Install (Recommended)

pip install clawteam

Option 2: Isolated User Environment

# Creates ~/.clawteam/.venv, links ~/.local/bin/clawteam
curl -sSL https://raw.githubusercontent.com/HKUDS/ClawTeam/main/scripts/install_clawteam.sh | bash

Option 3: Development Install from Source

git clone https://github.com/HKUDS/ClawTeam.git
cd ClawTeam
pip install -e .

Option 4: Local Checkout to Isolated Environment

# For contributing or testing modifications
scripts/clawteam_local_install

Optional: P2P Transport (ZeroMQ)

For low-latency inter-agent messaging across machines:

pip install -e ".[p2p]"

Post-Install Verification

clawteam --help
clawteam config health

The install scripts automatically detect and configure skills for supported agents. Check ~/.claude/skills/clawteam, ~/.codex/skills/clawteam, or equivalent directories for your agent.

Profile Configuration for Non-Default Providers

When using Claude Code through alternative providers (Moonshot Kimi, MiniMax, Gemini via Vertex), configure profiles instead of manual environment variables:

# List available provider templates
clawteam preset list

# Generate a reusable profile
clawteam preset generate-profile moonshot-cn claude --name claude-kimi

# Interactive setup wizard for first-time configuration
clawteam profile wizard

# Verify the profile works before spawning workers
MOONSHOT_API_KEY=sk-... clawteam profile test claude-kimi

Rules of thumb: preset = reusable provider template; profile = final runtime object used by spawn; wizard = easiest first-time path; doctor = fixes Claude Code onboarding state on fresh machines.

REAL Code Examples from the Repository

Example 1: Spawning a Specialized Research Agent

This is the fundamental operation that makes swarms possible—the leader creating workers with isolated environments:

# The leader agent runs this to create a worker with dedicated resources
clawteam spawn --team my-team \
  --agent-name worker1 \
  --task "Implement auth module"

What's happening under the hood: ClawTeam creates a new git worktree on branch clawteam/my-team/worker1, opens a tmux window with the agent's CLI, and auto-injects the coordination prompt. The worker receives its task context and immediately knows how to participate in team workflows. No Docker, no container orchestration, no cloud APIs—just filesystem isolation and process management.

For non-default providers, use your pre-configured profile:

clawteam spawn tmux --profile claude-kimi \
  --team my-team \
  --agent-name dave \
  --task "Refactor the auth flow"

Example 2: Worker Self-Coordination via Inbox

Workers don't need custom code to participate. The auto-injected coordination prompt teaches them standard CLI commands for team interaction:

# A worker agent checks its assigned tasks
clawteam task list my-team --owner me

# Worker reports completion back to leader
clawteam inbox send my-team leader \
  "Auth done. All tests passing."

Why this matters: This is framework-agnostic interoperability. Claude Code, Codex, nanobot, or your custom Python script—all use identical commands because ClawTeam operates at the shell level. The agent doesn't need to "know" about ClawTeam's API; it just executes shell commands like any other tool. This design decision eliminates integration friction entirely.

For broadcast announcements to all team members:

clawteam inbox broadcast my-team \
  "New priority: migrate to PostgreSQL 16 by EOD"

Example 3: Real-Time Swarm Monitoring

Observe your entire agent team simultaneously through multiple interfaces:

# Tiled tmux view: all agents visible in split windows
clawteam board attach my-team

# Web dashboard with real-time updates
clawteam board serve --port 8080

The attach command is transformative: Instead of switching between terminal tabs or losing context in chat histories, you see every agent's terminal simultaneously. Watch Claude Code architecting APIs while Codex writes tests while nanobot optimizes queries—all live, all visible, all coordinated.

For programmatic monitoring with JSON output:

# Auto-refreshing terminal dashboard every 3 seconds
clawteam board live my-team --interval 3 --json

Example 4: Complete Team Launch from Template

The most powerful one-liner in the repository—deploying an entire organizational structure:

# One command spawns 7 specialized investment analysts
clawteam launch hedge-fund \
  --team fund1 \
  --goal "Analyze AAPL, MSFT, NVDA for Q2 2026"

What executes automatically: The portfolio manager agent spawns, receives the goal, then calls clawteam spawn five times for analyst agents (Buffett value, growth disruption, technical indicators, fundamentals ratios, sentiment/news) and once for a risk manager. Each analyst works in isolation, submits signals to their inbox, and the risk manager consolidates before final portfolio decisions. The TOML template defines all roles, prompts, and communication patterns—no manual configuration per launch.

Example 5: Dependency-Managed Task Pipeline

Create sophisticated workflows where tasks unblock automatically:

# Create task with explicit dependency chain
clawteam task create my-team "Implement JWT auth" \
  -o backend1 \
  --blocked-by 1  # waits for task ID 1 (API schema design)

# When predecessor completes, dependent auto-unblocks
clawteam task update my-team 1 --status completed

# Block until entire pipeline finishes
clawteam task wait my-team --timeout 300

The dependency engine prevents coordination failures: Without explicit --blocked-by, agents might start work with incomplete prerequisites, wasting compute and generating conflicts. With it, the system maintains a consistent task state machine (pending → in_progress → completed/blocked) that agents query and update through simple CLI commands.

Advanced Usage & Best Practices

Profile-Driven Multi-Provider Orchestration

Don't hardcode API keys in environment variables. Use clawteam preset generate-profile to create reusable configurations for different providers, then spawn with --profile. This enables mixing agents from different vendors in one team—Claude via Moonshot for architecture, Codex via OpenAI for implementation, local models for testing—all coordinated through identical ClawTeam commands.

Workspace Checkpointing for Experiment Reproducibility

Before risky refactors, checkpoint agent state:

clawteam workspace checkpoint my-team experimenter-1
# Creates auto-commit on worktree branch
# Rollback if experiment fails: git reset --hard <checkpoint>

Lifecycle Management for Graceful Shutdowns

Agents request shutdown rather than dying unexpectedly:

# Worker signals completion
clawteam lifecycle request-shutdown my-team agent-3 --reason "Task complete"

# Leader approves, triggering cleanup
clawteam lifecycle approve-shutdown my-team 42 agent-3

Cross-Machine Distribution via Shared Filesystem

For GPU clusters, mount ~/.clawteam/ via NFS or SSHFS. All agents share identical team/task/inbox state without network configuration. For lower latency, enable P2P transport with automatic file fallback when network partitions occur.

Template Variable Substitution

Custom TOML templates support dynamic variables:

# In template: goal = "{goal}"
# At launch: --goal "Build X" substitutes into all agent prompts

Comparison with Alternatives

The decisive advantage: ClawTeam doesn't ask you to learn a new framework. It asks your agents to use tools they already understand—shell commands. Every other framework adds abstraction layers that humans must master. ClawTeam subtracts them, pushing complexity into the swarm's collective behavior rather than the operator's cognitive load.

FAQ

Q: Do I need cloud infrastructure or GPUs to use ClawTeam?

A: Absolutely not. ClawTeam runs on a single laptop with tmux and Python. The 8×H100 demo showcases scale, but the same clawteam spawn commands work on a MacBook Air. Start small, expand to clusters when needed.

Q: Which AI agents work with ClawTeam?

A: Any CLI agent executing shell commands: Claude Code, Codex, OpenClaw, nanobot, Kimi CLI, Cursor (experimental), or custom Python scripts. If it runs in terminal, it joins teams.

Q: How does ClawTeam prevent agents from conflicting?

A: Each agent receives an isolated git worktree on a dedicated branch. They literally cannot overwrite each other's files. The leader merges successful worktrees back to main using standard git operations.

Q: Is my code sent to external APIs?

A: Only through your chosen agent's normal operation. ClawTeam itself uses local filesystem storage exclusively. No telemetry, no cloud dependency, no data exfiltration by the framework.

Q: Can multiple humans share one agent team?

A: Yes. The multi-user workflow namespaces agents by user ID. Multiple developers can spawn agents in shared teams, with inbox routing ensuring messages reach correct recipients.

Q: What's the difference between file and p2p transport?

A: file (default) uses JSON files in ~/.clawteam/inboxes/—zero configuration, works anywhere with shared filesystem. p2p uses ZeroMQ PUSH/PULL sockets for lower latency, with automatic fallback to file transport if network fails.

Q: How do I debug when agents aren't coordinating?

A: Check clawteam board show <team> for task states, clawteam inbox peek <team> for undelivered messages, and clawteam workspace list <team> for worktree health. The --json flag on all commands enables programmatic debugging.

Conclusion

We've reached an inflection point in AI-assisted development. Solo agents—powerful as they are—hit complexity walls that no amount of prompt engineering overcomes. The future belongs to swarms: coordinated collectives that divide cognitive labor, cross-pollinate insights, and converge on solutions through emergent collaboration.

ClawTeam makes this future accessible today. Not through exotic infrastructure or proprietary cloud services, but through elegant simplicity: git worktrees for isolation, tmux for visibility, JSON files for state, and shell commands for universal interoperability. One pip install. One prompt to your leader agent. Then watch an entire organization of specialized workers self-assemble and execute.

The 2,430 experiments running autonomously across 8 GPUs aren't a demo—they're a preview of how every ambitious project will operate. The hedge fund analyzing markets while you sleep isn't a novelty—it's a template for any domain with repeatable team structures.

Stop coding alone. Stop manually coordinating agents. Stop letting infrastructure friction steal your focus.

Install ClawTeam. Spawn your first swarm. And discover what becomes possible when AI agents truly think and work as a team.

👉 Star ClawTeam on GitHub and join the swarm intelligence revolution.

8 agents × 8 H100s × 2,430 experiments × one CLI × one swarm. Your move. 🦞