Multi-Model Agent Teams for Claude Code — orchestrate 10 AI CLIs via Google A2A v0.3 + MCP bridge.

A Claude Code plugin that lets you build teams of AI CLIs and orchestrate them as A2A-compatible HTTP agents. Send the same question to 3 models and compare (fanout), chain write → review → synthesize across families (pipeline), or let each family spin up its own sub-sidecars for deeper reasoning (-team variants).

Every backend also supports a -team variant (e.g. kimi-code-team, claude-team) that internally creates 2-3 same-family sidecars via agent-teams, reasons independently, then synthesizes one final answer.

/plugin marketplace add LeoLin990405/mmteam-cc
/plugin install mmteam@mmteam-cc
/reload-plugins
/mmteam:setup

git clone https://github.com/LeoLin990405/mmteam-cc.git
cd mmteam-cc
bash install.sh          # copies bin to PATH + registers MCP
# Restart Claude Code to activate mcp__mmteam__a2a_* tools

/mmteam:create demo kimi:kimi-code gpt:codex gem:gemini

/mmteam:spawn demo               # headless (fastest)
/mmteam:spawn demo --dock         # cmux multi-pane grid
/mmteam:spawn demo --dock --monitor  # + top event strip

# Single smart-routed ask
/mmteam:ask demo "用 Python 实现 LRU Cache"

# Parallel fanout — 3 models answer, compare consensus
/mmteam:fanout demo "Is 91 prime? Answer ONLY prime or composite" --agents kimi,gpt,gem

# Sequential pipeline — write → review → synthesize
/mmteam:pipeline demo "Build a REST API for todo app" --writer kimi --reviewer gem --synth gpt

# Direct send to one teammate
/mmteam:send demo kimi "分析这段 500K token 日志"

# Watch live output in cmux dashboard
/mmteam:watch demo

/mmteam:stop demo      # stop daemons, keep data
/mmteam:destroy demo   # remove everything

After install, Claude Code can call mmteam natively via 20+ MCP tools:

mcp__mmteam__a2a_spawn     mcp__mmteam__a2a_send
mcp__mmteam__a2a_fanout    mcp__mmteam__a2a_pipeline
mcp__mmteam__a2a_ask       mcp__mmteam__a2a_stop
mcp__mmteam__a2a_ls        mcp__mmteam__a2a_watch
...

This means Claude can autonomously decide to dispatch tasks to teammates during a conversation — no slash command needed.

Claude Code (main session)
  │
  ├── /mmteam:* slash commands
  │      └── mmteam-bridge.mjs → bin/mmteam CLI
  │
  └── mcp__mmteam__a2a_* (native MCP tools)
         └── bin/mmteam-mcp.py (stdio JSON-RPC)
                │
                │  JSON-RPC 2.0 over HTTP + Bearer token
                │
       ┌────────┼────────┬────────────┐
       ▼        ▼        ▼            ▼
    a2a-srv  a2a-srv  a2a-srv     a2a-srv
    :port₁   :port₂   :port₃      :portₙ
       │        │        │            │
       ▼        ▼        ▼            ▼
    kimi-code  codex   gemini    glm-code-team
    (headless) (headless) (headless) (agent-teams)
                                      └→ 2-3 GLM sidecars
                                         → synthesized reply

Protocol: Google A2A v0.3 minimal subset — Agent Card discovery, message/send, tasks/get, tasks/cancel, Bearer auth.

Append -team to any CLI name to enable internal multi-sidecar reasoning:

/mmteam:create deep kt:kimi-code-team ct:claude-team
/mmteam:spawn deep
/mmteam:fanout deep "Design a distributed cache" --agents kt,ct

Each -team teammate internally:

1. Sets CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1
2. Wraps the prompt with TeamCreate instructions
3. Spawns 2-3 same-family sidecars that reason independently
4. Synthesizes a single final answer

Agent Card shows family: anthropic-agent-teams and skills[0]: parallel-sidecars.

Register a teammate running on another machine:

# On remote host (Mac Studio):
mmteam a2a spawn myteam
jq -r '.kimi.bearer_token' ~/.claude/teams/myteam/a2a-registry.json

# On local host:
/mmteam:remote myteam kimi-studio http://studio.local:55723/ --token <TOKEN>
/mmteam:send myteam kimi-studio "Analyze this large codebase"

See docs/REMOTE-TEAMMATE.md.

Run bash doctor.sh to check your setup.

cn-cc provides single-point /cn:* slash commands for routing individual tasks to Chinese model backends.

mmteam-cc provides multi-model orchestration — teams, fanout, pipelines, consensus, cross-host.

They are complementary:

• Simple one-off Chinese model task → /cn:ask
• Multi-model consensus or staged workflow → /mmteam:fanout or /mmteam:pipeline
• Both can be installed side by side.

Full diagnostics: bash doctor.sh

Local project gate: npm run check runs JavaScript syntax checks, Python bytecode compilation, shell syntax checks, and smoke-script syntax checks.

mmteam-cc/
├── bin/                    # Vendored mmteam scripts
│   ├── mmteam              # Main CLI
│   ├── mmteam-a2a-server.py # Per-teammate HTTP daemon
│   ├── mmteam-a2a-monitor.py # Event stream viewer
│   └── mmteam-mcp.py       # MCP stdio server (20 tools)
├── plugins/mmteam/         # Claude Code plugin
│   ├── commands/           # 15 slash commands
│   ├── agents/             # mmteam-orchestrator
│   ├── skills/mmteam/      # SKILL.md + ROUTING.md
│   └── scripts/            # Bridge + MCP registrar
├── install.sh              # One-shot bootstrap
├── uninstall.sh            # Clean removal
├── doctor.sh               # Health diagnostics
├── docs/                   # Deep-dive documentation
├── examples/               # Usage recipes
└── tests/smoke/            # CI smoke tests

Apache License 2.0 — see LICENSE.

This project implements wire-compatible subsets of the A2A Protocol and MCP but is not affiliated with their respective organizations. See NOTICE.