convez.ai
MCP ServerAutomaçãoAvançado
timzen

MPT MCP Server

timzen/mpt-mcp-server

Coordenação multi-agente e gestão de tasks via MCP protocol

Ver no GitHub
InicialPluginScore · de 0 a 100

Servidor MCP que expõe ferramentas de colaboração em equipe para fluxos de trabalho multi-agente. Funciona como alternativa agnóstica a harnesses como Claude Code, Cursor, Kiro e Codex, oferecendo coordenação de equipe via daemon com suporte a memory persistente, gerenciamento de stories/tasks e orquestração de agentes em tmux.

Instalação

$npx convez-ai timzen/mpt-mcp-server

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Orquestração de múltiplos agentes de IA em workflows colaborativos
  • Persistência e busca de conhecimento compartilhado entre agentes
  • Gerenciamento de histórico de tasks, stories e decisões de equipe
  • Automação de pipelines com múltiplos harnesses (Claude, Cursor, Kiro, Codex)

Veredito

Projeto WIP em fase inicial com arquitetura interessante para multi-agência, mas ainda carece de documentação estável e casos de uso consolidados. Recomendado para quem busca experimentar coordenação de agentes em escala — não para produção imediata.

README

A tradução para o português deste README ainda está na fila.

mpt-mcp-server

An MCP (Model Context Protocol) server that exposes team collaboration tools for multi-agent workflows. Compatible with Claude Code, Cursor, Kiro, and Codex via their MCP configuration.

This is the harness-agnostic alternative to pi-pizza-team. Where pi-pizza-team integrates directly into Pi's extension system (TUI widgets, permission hooks, pi.sendUserMessage), mpt-mcp-server provides the same daemon-backed team coordination as a standalone MCP tool layer that any harness can use. Same daemon, same agent protocol, same workflow model — just without the Pi-specific integration points.

Tools

ToolDescription
save_memoryPersist knowledge, decisions, or conventions to team memory (via daemon)
search_memorySearch team memory by keyword, with optional category filter (via daemon)
upload_attachmentUpload a file artifact to a task (requires taskId)
create_storyCreate a new story (unit of plannable work with tasks)
edit_storyUpdate an existing story's details, status, or dependencies
add_taskAdd a task to a story for teammates to execute
team_statusGet a summary of stories, tasks, members, and inbox
queue_requestQueue an async request for the assistant to process
spawn_agentSpawn a teammate agent in a tmux window (leader can observe/interact)
dismiss_agentStop a running teammate and close its tmux window
list_agentsList all active teammate agents and their tmux status
get_next_workPoll daemon for next task with teammate-allowed transitions
claim_taskClaim a task and start working (daemon transitions to working state)
release_taskRelease a task after work (daemon advances to next state)
post_commentPost a comment on a task (status updates, questions)
report_token_usageReport token usage (input/output tokens) for a task

Setup

npm install

Usage

Start the MCP server (stdio transport)

npm start
# or
node src/index.mjs

Start the Task Runner

The task runner is a persistent loop that polls a daemon for tasks, executes them via the claude CLI, and reports results back.

npm run runner
# or
node src/runners/claude/runner.mjs

Start the Task Runner (Kiro)

The Kiro runner uses kiro-cli chat --no-interactive with MCP support.

npm run runner:kiro
# or
node src/runners/kiro/runner.mjs

Start the Task Runner (Codex)

The Codex runner uses the same shared loop as Claude/Kiro but without MCP tools (codex doesn't support MCP).

npm run runner:codex
# or
node src/runners/codex/runner.mjs

Leader Mode (tmux-based team orchestration)

The recommended way to use mpt is as a leader + teammates setup. One harness (e.g., Pi or Claude Code) runs as the leader with the MCP server. It can spawn teammate agents into tmux windows that users can hop into to observe or course-correct.

# 1. Start a tmux session (or let mpt create one automatically)
tmux new-session -s mpt-team

# 2. Run your leader harness with the MCP server configured
#    The leader can now call spawn_agent to create teammates

# 3. Teammates appear as new tmux windows in the session
#    Hop to any window: tmux select-window -t mpt-team:<agent-name>

How it works:

  • The leader's MCP server instance has spawn_agent, dismiss_agent, and list_agents tools
  • spawn_agent creates a new tmux window, launches the harness with its own MCP config
  • Each teammate gets its own MCP server instance (stdio is 1:1)
  • Users can switch to any teammate's window to pair, observe, or intervene
  • The daemon remains the shared state layer for task coordination

Environment:

VariableDefaultDescription
MPT_TMUX_SESSIONmpt-teamtmux session name for all agent windows

Environment Variables

VariableDefaultDescription
MPT_DAEMON_URL(required)Daemon HTTP endpoint (my-pizza-team)
MPT_AGENT_IDmpt-agent-<pid>This agent's identifier
MPT_ROLEleaderRole for tool filtering: leader, teammate, or assistant
MPT_POLL_INTERVAL5Seconds between polls when idle
MPT_WORK_DIRCurrent directoryWorking directory for claude sessions
MPT_MCP_SERVERAuto-detectedPath to mpt-mcp-server entry point
MPT_CLAUDE_BINclaudePath to claude CLI binary (claude runner)
MPT_KIRO_BINkiro-cliPath to kiro-cli binary (kiro runner)
MPT_CODEX_BINcodexPath to codex CLI binary (codex runner)
MPT_MAX_RETRIES3Max retries for reporting results
MPT_LOG_LEVELinfoLog level (debug, info, warn, error)
MPT_TMUX_SESSIONmpt-teamtmux session name for spawned agents

Spawn an Agent (mpt spawn)

Launch any configured harness with a prompt:

# Use default harness (claude-code)
node src/cli/spawn.mjs --prompt="Fix the bug in main.js"

# Specify harness
node src/cli/spawn.mjs --harness=pi --prompt="Refactor the auth module"
node src/cli/spawn.mjs --harness=codex --prompt="Add tests for utils"

# List available harnesses
node src/cli/spawn.mjs --list

Project Config (mpt.config.json)

Optional config file in the project root for customizing harnesses:

{
  "daemonUrl": "http://localhost:3100",
  "defaults": {
    "harness": "claude-code",
    "pollInterval": 5
  },
  "harnesses": {
    "claude-code": {
      "command": "/custom/path/to/claude",
      "env": { "CUSTOM_VAR": "value" }
    },
    "my-custom-agent": {
      "name": "my-custom-agent",
      "description": "My custom agent",
      "command": "my-agent",
      "args": ["--run", "{{prompt}}"],
      "capabilities": { "mcp": false, "midTaskComm": false, "fileEdit": true, "shell": true, "streaming": false },
      "env": {},
      "workDir": "{{workDir}}"
    }
  }
}

Template variables: {{workDir}}, {{mcpConfigPath}}, {{mcpServerPath}}, {{prompt}}, {{taskId}}, {{agentId}}, {{agentsmd}}

Role-Based Tool Filtering

Each MCP server instance exposes tools based on MPT_ROLE:

ToolLeaderTeammateAssistant
create_story✅❌✅
edit_story✅❌✅
add_task✅❌✅
team_status✅❌❌
queue_request✅❌✅
save_memory✅❌✅
search_memory✅✅✅
upload_attachment❌✅❌
get_next_work❌✅❌
claim_task❌✅❌
release_task❌✅❌
post_comment✅✅❌
report_token_usage❌✅❌
spawn_agent✅❌❌
dismiss_agent✅❌❌
list_agents✅❌❌

Spawned teammates automatically get MPT_ROLE=teammate. The leader's own MCP instance defaults to MPT_ROLE=leader.

Configure in Claude Code

Add to your MCP config (~/.config/claude/mcp.json or project .mcp.json):

{
  "mcpServers": {
    "mpt": {
      "command": "node",
      "args": ["/path/to/mpt-mcp-server/src/index.mjs"]
    }
  }
}

Configure in Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "mpt": {
      "command": "node",
      "args": ["/path/to/mpt-mcp-server/src/index.mjs"]
    }
  }
}

Configure in Kiro

Add to your Kiro MCP settings with the same command/args pattern.

Directory Structure

mpt-mcp-server/
├── README.md                 # This file
├── AGENTS.md                 # Instructions for AI agents
├── package.json              # Package manifest
├── docs/
│   ├── ARCHITECTURE.md       # Internal technical docs
│   └── DESIGN.md             # Design philosophy
├── src/
│   ├── index.mjs             # Entry point — starts stdio MCP server
│   ├── daemon-client.mjs     # HTTP client for daemon API
│   ├── tmux.mjs              # tmux session/window management helpers
│   ├── config/
│   │   ├── index.mjs         # Config public API
│   │   ├── harnesses.mjs     # Default harness definitions (pi, claude-code, codex)
│   │   └── loader.mjs        # Config file loader + merger
│   ├── cli/
│   │   └── spawn.mjs         # `mpt spawn --harness=X` command
│   ├── tools/
│   │   ├── registry.mjs      # Tool registry (list + dispatch)
│   │   ├── save_memory.mjs   # save_memory tool
│   │   ├── search_memory.mjs # search_memory tool
│   │   ├── upload_attachment.mjs
│   │   ├── create_story.mjs  # Create a story (leader)
│   │   ├── edit_story.mjs    # Update a story (leader)
│   │   ├── add_task.mjs      # Add task to story (leader)
│   │   ├── team_status.mjs   # Get team status summary
│   │   ├── queue_request.mjs # Queue async assistant request
│   │   ├── report_token_usage.mjs # Report token usage per task
│   │   ├── spawn_agent.mjs   # Spawn teammate in tmux window
│   │   ├── dismiss_agent.mjs # Stop teammate, close tmux window
│   │   ├── list_agents.mjs   # List active teammates
│   │   ├── get_next_work.mjs # Poll daemon for next task (workflow-aware)
│   │   ├── claim_task.mjs    # Claim task ownership
│   │   ├── release_task.mjs  # Release task back to pool
│   │   └── post_comment.mjs  # Post comment on a task
│   ├── runners/
│   │   ├── shared/
│   │   │   ├── loop.mjs      # Shared runner loop (register/poll/claim/transition/execute/release)
│   │   │   ├── config.mjs    # Common env var config loader
│   │   │   └── logger.mjs    # Structured JSON logger
│   │   ├── claude/
│   │   │   ├── runner.mjs    # Entry point (thin: imports shared loop + adapter)
│   │   │   └── adapter.mjs   # Claude-specific CLI wrapper (claude --print)
│   │   ├── kiro/
│   │   │   ├── runner.mjs    # Entry point (thin: imports shared loop + adapter)
│   │   │   └── adapter.mjs   # Kiro-specific CLI wrapper (kiro-cli chat)
│   │   └── codex/
│   │       ├── runner.mjs    # Entry point (thin: imports shared loop + adapter)
│   │       ├── adapter.mjs   # Codex-specific CLI wrapper (codex --auto-edit)
│   │       └── codex-runner.sh  # Legacy bash runner (deprecated)
└── tests/
    ├── tools.test.mjs        # MCP tool integration tests
    ├── tmux-tools.test.mjs   # tmux spawn/dismiss/list tool tests
    ├── shared-runner.test.mjs # Shared loop, adapters, and utilities
    ├── runner.test.mjs       # Claude runner component tests (legacy)
    ├── kiro-runner.test.mjs  # Kiro runner tests
    ├── codex-runner.test.sh  # Codex runner tests (legacy bash)
    ├── config.test.mjs       # Harness config tests
    ├── spawn.test.mjs        # Spawn CLI tests
    ├── integration.test.mjs  # Multi-harness integration tests
    └── helpers/
        ├── mock-daemon.mjs   # Mock HTTP daemon for tests
        └── mock-agents.mjs   # Simulated agent adapters

Testing

npm test

Requirements

  • Node.js >= 20.0.0
  • @modelcontextprotocol/sdk (installed via npm)
Ver o original no GitHub

Bom para

Relacionados