ClawTeam: the CLI whose users are the agents themselves

Agents as the framework's user class

ClawTeam's inversion is stated in the first row of its own comparison table: the users of this tool are the AI agents themselves, where other multi-agent frameworks serve humans writing orchestration code. The intended loop is that a human provides the goal, and an agent team orchestrates everything else, spawning swarms, delegating tasks and delivering results from one command line. Compatibility is correspondingly broad rather than framework-deep: Claude Code, Codex, OpenClaw, nanobot and Cursor are all named as working, along with any CLI agent, because the coordination surface is the shell. The project comes from HKUDS, a university data science lab, is MIT licensed, and carries README translations in Chinese and Korean alongside the English one, an internationalism that matches its subject matter. Its own news section records a public launch on 2026-03-18, which makes this a young tool with an unusually clear thesis about who its users are.

spawn: a worktree, a tmux window, an identity

Team formation is one command issued by the leader agent, and the per-worker provisioning is the design decision that matters most:

# The leader agent runs:
clawteam spawn --team my-team \
  --agent-name worker1 \
  --task "Implement auth module"

Each spawned worker automatically gets its own git worktree, its own tmux window and its own identity. The worktree choice is the interesting one, and the README argues it directly: isolation happens in real branches with real diffs, rather than in containers or virtual environments, so the output of parallel agents is reviewable with the tools every developer already uses. The task string arrives with the spawn, so a worker starts with its assignment rather than discovering it, and the tmux window means every agent's activity is visible as a terminal pane from the first moment.

task list and inbox send: coordination as CLI

The inter-agent protocol is two commands wide. A worker checks its assignments and reports back:

# A worker agent checks tasks:
clawteam task list my-team --owner me
# Then reports back:
clawteam inbox send my-team leader \
  "Auth done. All tests passing."

These commands are auto-injected into agent prompts, so an agent does not need ClawTeam-specific training to participate, only the ability to run commands and read output. Task allocation includes smart dependency management, workers update task status as they go, and results flow to the leader's inbox. The infrastructure claim follows from this minimalism: just a filesystem and tmux, no Redis, message queues or databases, which removes both the deployment weight and the operational failure surface that heavier frameworks carry.

board attach and board serve: watching instead of driving

The human's window onto the swarm is deliberately observational:

# Watch all agents simultaneously
clawteam board attach my-team
# Or open the web dashboard
clawteam board serve --port 8080

board attach tiles every agent's tmux pane into one view, and board serve opens a web dashboard for the same information in a browser. The division of labor is explicit: the leader handles coordination, monitors team performance, identifies bottlenecks and reallocates resources or redirects effort dynamically, and the human intervenes only when they want to. Watching a swarm work is also the debugging story, since a stalled agent is a pane that stopped changing rather than a log message to correlate, an operational property terminal users will recognize as genuinely useful. Both views render the same team state from different angles, so the web dashboard is a convenience rather than a dependency, and everything it shows is derivable from the tmux panes it summarizes.

The comparison its authors draw against frameworks

The README's comparison table is the alternatives section in miniature, and its rows name real differences in approach rather than benchmark scores. Setup is a pip install plus one prompt to the leader, against Docker, cloud APIs and YAML configuration for others. Infrastructure is filesystem and tmux against Redis, message queues and databases. Agent support is any CLI agent against framework-specific runtimes. Isolation is git worktrees against containers or virtual environments. And intelligence is the swarm self-organizing through CLI commands against hard-coded orchestration logic. The honest reading is that ClawTeam trades the guarantees of a managed runtime, central state, typed message channels, for frictionless adoption and universal agent compatibility, a trade that favors small teams of CLI agents over large supervised fleets.

Eight agents, eight H100s, karpathy's autoresearch

The flagship demonstration is autonomous machine learning research: eight specialized sub-agents across eight H100 GPUs, with the leader autonomously designing experiments and dynamically reallocating resources based on real-time performance, an arrangement the project credits as based on karpathy's autoresearch. The broader use-case families show the intended range: AI research automation including hypothesis generation and validation and self-improving architectures, agentic engineering including autonomous full-stack development and collaborative open source work, an AI hedge fund pattern spanning market research, portfolio optimization, risk assessment and algorithmic execution, and finally custom swarms, research teams, investment committees, business operations and content studios. The breadth is aspirational marketing in part, but the mechanism is the same in every case, and the ML demonstration is the one with published video evidence in the README.

v0.3.0, an MCP server and a skills lockfile

Maturity signals are mixed and visible. The classifier says Alpha, the last push was 2026-05-09, and releases run only to v0.2.0, the stabilization release of 2026-03-23, five days after the first PyPI release, while pyproject already declares 0.3.0. The declared baseline covers config management, multi-user workflows, a web UI, P2P transport via a pyzmq extra and team templates, with Redis available as another optional extra. Two entry points ship: clawteam for the CLI, built on typer with pydantic models and rich output, and clawteam-mcp, an MCP server so LLM hosts can drive teams through the Model Context Protocol rather than the shell. A skills directory with a committed skills-lock.json mirrors the agent-skill packaging convention, a ROADMAP.md states direction, and the website is a small React and Vite application. Python support runs 3.10 through 3.12.

Editorial conclusion

Use ClawTeam when you already drive coding or research through CLI agents and want a team of them to spawn, divide and report work without you writing orchestration code. Choose a conventional multi-agent framework when you want to program the orchestration yourself or need its message-queue infrastructure. Verify first that your agents can execute shell commands, since the whole coordination protocol is CLI-based, check the Alpha maturity status against your production appetite, and note the last push was 2026-05-09 before building on the current release.

Frequently asked questions

Official sources

1. HKUDS/ClawTeam on GitHub
2. Issues
3. License: MIT
4. README
5. Releases