/mpm - Claude MPM Framework Guide
What is Claude MPM?
Claude MPM (Multi-Agent Project Manager) extends Claude Code with multi-agent orchestration, project-specific PM instructions, persistent agent memory, real-time monitoring via a WebSocket dashboard, and automation hooks. A PM agent coordinates specialized agents to complete work through structured delegation using the Task tool. MPM manages skills deployment, agent selection, session continuity, and cross-project messaging.
How Delegation Works
The PM receives user requests and delegates to specialized agents via the Task tool:
Task(description="[what to do]", subagent_type="[agent-type]")
- PM analyzes the request, breaks it into subtasks, and selects the right agent for each
- Each agent focuses on its specialty and returns results to PM
- Parallel execution when tasks are independent (e.g., frontend + backend simultaneously)
- Sequential execution when tasks depend on each other (e.g., implement, then test, then deploy)
- Agents should focus on their delegated task and return clear results -- PM handles orchestration
- If an agent needs help from another specialty, it reports back to PM rather than self-delegating
Typical workflow: Research -> Engineer -> Ops (deploy) -> Ops (verify) -> QA -> Documentation
Available Agents
Core Agents (always deployed)
| Agent |
subagent_type |
Role |
| Engineer |
engineer |
Code implementation, refactoring, debugging |
| Research |
Research |
Investigation, analysis, codebase exploration |
| QA |
qa |
Testing, validation, quality assurance |
| Documentation |
documentation |
Docs generation, API specs, guides |
| Ops |
ops |
Deployment, infrastructure, operations |
| Security |
security |
Vulnerability assessment, security review |
| Ticketing |
ticketing |
Ticket tracking and workflow management |
Extended Agents (deployed per project)
| Agent |
subagent_type |
Role |
| Version Control |
version-control |
Git operations, PR management, tagging |
| Code Analyzer |
code-analyzer |
Codebase analysis, architecture review |
| Code Review |
code-review |
PR review, semantic analysis |
| Data Engineer |
data-engineer |
Database management, API integrations |
| Product Owner |
product-owner |
Requirements, user stories, prioritization |
Language-Specific Engineers
Deployed automatically based on project toolchain detection:
python-engineer, typescript-engineer, javascript-engineer, golang-engineer, rust-engineer, java-engineer, ruby-engineer, php-engineer, dart-engineer
Framework specialists: nextjs-engineer, react-engineer, svelte-engineer, tauri-engineer, phoenix-engineer
Platform Ops
local-ops (local dev, PM2, Docker), vercel-ops, gcp-ops, digitalocean-ops, clerk-ops, railway-ops
Specialized Agents
api-qa, web-qa, code-review, code-analyzer, refactoring-engineer, prompt-engineer, content-agent, imagemagick, memory-manager
Skills System
Skills are markdown files deployed to .claude/skills/{name}/SKILL.md that provide agents with specialized knowledge and procedures.
- User-invocable skills respond to
/skill-name slash commands (e.g., /mpm, /mpm-help)
- Non-invocable skills activate automatically based on context triggers
- Skills have frontmatter with:
name, description, user-invocable, version, category, tags
- Categories:
mpm-command, pm-workflow, pm-reference, toolchains-*, universal-*
- Progressive disclosure: Metadata loads first; full content loads on trigger
- Skills can include
references/ subdirectories for detailed supplementary content
Memory System
Agent memories provide persistent context across sessions.
- Storage:
.claude-mpm/memories/{agent_id}_memories.md
- PM manages memory files directly (read, consolidate, save)
- Each agent has domain-specific memory categories
- Trigger phrases: "remember", "don't forget", "always", "never", "going forward"
- Size limit: 80KB per file (~20k tokens)
- Routing: Keyword-based routing sends memories to the appropriate agent
- Memories are automatically loaded when an agent is delegated work
Hooks and Dashboard
Claude Code hooks capture tool usage, responses, and agent activity in real-time.
Event flow: Hook -> Connection Manager -> Monitor Server -> Dashboard
Dashboard at http://localhost:8765/ shows:
- Tools being used and files being read/written
- Agent delegations and completions
- Session timeline and activity feed
Monitor commands: /mpm-monitor start, /mpm-monitor stop, /mpm-monitor status
Available Commands
| Command |
Description |
/mpm |
This guide -- MPM overview and framework reference |
/mpm-help |
Detailed help for MPM commands |
/mpm-init |
Initialize or update a project for MPM |
/mpm-status |
System health and status |
/mpm-doctor |
Run diagnostic checks |
/mpm-config |
Manage configuration |
/mpm-monitor |
Control monitoring server and dashboard |
/mpm-version |
Version information |
/mpm-message |
Cross-project messaging |
/mpm-organize |
Intelligent file consolidation |
/mpm-ticket-view |
Ticketing workflow management |
/mpm-session-pause |
Save session state for later |
/mpm-session-resume |
Resume from paused session |
/mpm-postmortem |
Analyze session errors |
Cross-Project Messaging
/mpm-message sends asynchronous messages between Claude MPM instances running in different projects on the same machine.
- Messages stored in a shared SQLite database at
~/.claude-mpm/messaging.db
- Checked periodically: on session start, every 10 commands, every 30 minutes
- Message types:
task, request, notification, reply
- Always use the
MessageService API or claude-mpm message CLI -- never query the database directly
For Agents: How to Work Within MPM
When you receive a delegated task from PM:
- Focus on your delegated task -- PM handles orchestration and coordination
- Return clear results with evidence: file paths created/modified, test counts and pass rates, URLs, error details
- Your agent memory is loaded automatically -- reference it for project-specific conventions and past decisions
- If you need another specialty, report back to PM with what you need rather than self-delegating
- Track files you create or modify -- PM will handle git operations and file tracking
- Follow project patterns -- check your memory and existing code conventions before implementing
- Escalate blockers immediately -- do not silently fail or produce partial results without explanation
1---2name: mpm-33description: Access Claude MPM functionality and manage multi-agent orchestration4user-invocable: true5version: "2.0.0"6category: mpm-command7tags: [mpm-command, system, pm-required, framework]8metadata:9 source: https://skillmd.com/skills/bobmatnyc/mpm-310---1112# /mpm - Claude MPM Framework Guide1314## What is Claude MPM?1516Claude MPM (Multi-Agent Project Manager) extends Claude Code with multi-agent orchestration, project-specific PM instructions, persistent agent memory, real-time monitoring via a WebSocket dashboard, and automation hooks. A **PM agent** coordinates specialized agents to complete work through structured delegation using the Task tool. MPM manages skills deployment, agent selection, session continuity, and cross-project messaging.1718## How Delegation Works1920The PM receives user requests and delegates to specialized agents via the **Task tool**:2122```23Task(description="[what to do]", subagent_type="[agent-type]")24```2526- PM analyzes the request, breaks it into subtasks, and selects the right agent for each27- Each agent focuses on its specialty and returns results to PM28- **Parallel execution** when tasks are independent (e.g., frontend + backend simultaneously)29- **Sequential execution** when tasks depend on each other (e.g., implement, then test, then deploy)30- Agents should focus on their delegated task and return clear results -- PM handles orchestration31- If an agent needs help from another specialty, it reports back to PM rather than self-delegating3233**Typical workflow**: Research -> Engineer -> Ops (deploy) -> Ops (verify) -> QA -> Documentation3435## Available Agents3637### Core Agents (always deployed)3839| Agent | subagent_type | Role |40|-------|---------------|------|41| Engineer | `engineer` | Code implementation, refactoring, debugging |42| Research | `Research` | Investigation, analysis, codebase exploration |43| QA | `qa` | Testing, validation, quality assurance |44| Documentation | `documentation` | Docs generation, API specs, guides |45| Ops | `ops` | Deployment, infrastructure, operations |46| Security | `security` | Vulnerability assessment, security review |47| Ticketing | `ticketing` | Ticket tracking and workflow management |4849### Extended Agents (deployed per project)5051| Agent | subagent_type | Role |52|-------|---------------|------|53| Version Control | `version-control` | Git operations, PR management, tagging |54| Code Analyzer | `code-analyzer` | Codebase analysis, architecture review |55| Code Review | `code-review` | PR review, semantic analysis |56| Data Engineer | `data-engineer` | Database management, API integrations |57| Product Owner | `product-owner` | Requirements, user stories, prioritization |5859### Language-Specific Engineers6061Deployed automatically based on project toolchain detection:6263`python-engineer`, `typescript-engineer`, `javascript-engineer`, `golang-engineer`, `rust-engineer`, `java-engineer`, `ruby-engineer`, `php-engineer`, `dart-engineer`6465**Framework specialists**: `nextjs-engineer`, `react-engineer`, `svelte-engineer`, `tauri-engineer`, `phoenix-engineer`6667### Platform Ops6869`local-ops` (local dev, PM2, Docker), `vercel-ops`, `gcp-ops`, `digitalocean-ops`, `clerk-ops`, `railway-ops`7071### Specialized Agents7273`api-qa`, `web-qa`, `code-review`, `code-analyzer`, `refactoring-engineer`, `prompt-engineer`, `content-agent`, `imagemagick`, `memory-manager`7475## Skills System7677**Skills** are markdown files deployed to `.claude/skills/{name}/SKILL.md` that provide agents with specialized knowledge and procedures.7879- **User-invocable skills** respond to `/skill-name` slash commands (e.g., `/mpm`, `/mpm-help`)80- **Non-invocable skills** activate automatically based on context triggers81- Skills have **frontmatter** with: `name`, `description`, `user-invocable`, `version`, `category`, `tags`82- **Categories**: `mpm-command`, `pm-workflow`, `pm-reference`, `toolchains-*`, `universal-*`83- **Progressive disclosure**: Metadata loads first; full content loads on trigger84- Skills can include `references/` subdirectories for detailed supplementary content8586## Memory System8788Agent memories provide persistent context across sessions.8990- **Storage**: `.claude-mpm/memories/{agent_id}_memories.md`91- PM manages memory files directly (read, consolidate, save)92- Each agent has domain-specific memory categories93- **Trigger phrases**: "remember", "don't forget", "always", "never", "going forward"94- **Size limit**: 80KB per file (~20k tokens)95- **Routing**: Keyword-based routing sends memories to the appropriate agent96- Memories are automatically loaded when an agent is delegated work9798## Hooks and Dashboard99100Claude Code hooks capture tool usage, responses, and agent activity in real-time.101102**Event flow**: Hook -> Connection Manager -> Monitor Server -> Dashboard103104**Dashboard** at `http://localhost:8765/` shows:105- Tools being used and files being read/written106- Agent delegations and completions107- Session timeline and activity feed108109**Monitor commands**: `/mpm-monitor start`, `/mpm-monitor stop`, `/mpm-monitor status`110111## Available Commands112113| Command | Description |114|---------|-------------|115| `/mpm` | This guide -- MPM overview and framework reference |116| `/mpm-help` | Detailed help for MPM commands |117| `/mpm-init` | Initialize or update a project for MPM |118| `/mpm-status` | System health and status |119| `/mpm-doctor` | Run diagnostic checks |120| `/mpm-config` | Manage configuration |121| `/mpm-monitor` | Control monitoring server and dashboard |122| `/mpm-version` | Version information |123| `/mpm-message` | Cross-project messaging |124| `/mpm-organize` | Intelligent file consolidation |125| `/mpm-ticket-view` | Ticketing workflow management |126| `/mpm-session-pause` | Save session state for later |127| `/mpm-session-resume` | Resume from paused session |128| `/mpm-postmortem` | Analyze session errors |129130## Cross-Project Messaging131132`/mpm-message` sends asynchronous messages between Claude MPM instances running in different projects on the same machine.133134- Messages stored in a shared SQLite database at `~/.claude-mpm/messaging.db`135- Checked periodically: on session start, every 10 commands, every 30 minutes136- Message types: `task`, `request`, `notification`, `reply`137- Always use the `MessageService` API or `claude-mpm message` CLI -- never query the database directly138139## For Agents: How to Work Within MPM140141When you receive a delegated task from PM:1421431. **Focus on your delegated task** -- PM handles orchestration and coordination1442. **Return clear results with evidence**: file paths created/modified, test counts and pass rates, URLs, error details1453. **Your agent memory is loaded automatically** -- reference it for project-specific conventions and past decisions1464. **If you need another specialty**, report back to PM with what you need rather than self-delegating1475. **Track files you create or modify** -- PM will handle git operations and file tracking1486. **Follow project patterns** -- check your memory and existing code conventions before implementing1497. **Escalate blockers immediately** -- do not silently fail or produce partial results without explanation