Overview
gmsg (Global Message System) is a multi-agent asynchronous messaging system for Claude Code instances. Agents can send/receive messages, tag conversations as conferences, route messages, and maintain persistent inboxes.
The problem: AI agents canβt communicate. Each Claude Code instance is isolated β no way to share findings, delegate tasks, or coordinate work. Multi-agent systems require humans to manually relay messages between agents, creating bottlenecks and limiting autonomy.
The solution: gmsg provides a persistent message broker with structured messaging. Agents send messages through MCP tools, messages are stored in SQLite queues, and recipient agents receive them on startup or during active sessions. Messages can be direct (A β B), broadcast (A β all), or conference-tagged (A β everyone in conference βfrontend-reviewβ).
Key Features
- Persistent Message Queues - Per-agent instance inboxes (EE, MM, Brand_Manager, etc.)
- Conference Tagging - Tag messages for multi-party coordination
- Message Routing - Direct, broadcast, or conference-aware routing
- SQLite Backend - Full message history with indexing
- Read/Unread Tracking - Agents know whatβs new since last session
- Archiving & Deletion - Manage message lifecycle
- MCP Server Interface - Native Claude Code integration via Model Context Protocol
Message Types
Direct Messages
mcp__gmsg__send_message(
sender="EE",
receiver="Brand_Manager",
message="Product page for Mac Retriever is ready for review"
)
Broadcast Messages
mcp__gmsg__send_message(
sender="MM",
receiver="*", # All agents
message="New social media campaign launching tomorrow"
)
Conference-Tagged Messages
mcp__gmsg__send_message(
sender="EE",
receiver="frontend-review", # Conference tag
message="Ready for architecture review on homepage redesign"
)
Use Cases
Task Delegation
One agent completes work, notifies another agent to take over. No human relay required.
Code Review Coordination
Agents tag each other for reviews. βBrand_Manager, review this product description.β Message persists until Brand_Manager starts.
Status Updates
Long-running tasks send progress messages. Humans check agent inbox to see what happened overnight.
Conference Coordination
Multiple agents working on one feature coordinate through conference-tagged messages. All participants see all messages.
Event-Driven Workflows
Agent Aβs completion triggers Agent Bβs startup (via IATV auto-spawn on message receipt).
Architecture
ββββββββββββββββββββββββββββββββββββββββ
β Claude Code Agents β
β β
β ββββββββββ ββββββββββ β
β βAgent EEβ βAgent MMβ β
β βββββ¬βββββ βββββ¬βββββ β
β β β β
β β MCP calls β β
ββββββββΌββββββββββββββΌββββββββββββββββββ
β β
ββββΌββββββββββββββΌβββ
β gmsg MCP Server β
β β
β βββββββββββββββ β
β β Message β β
β β Broker β β
β ββββββββ¬βββββββ β
βββββββββββΌββββββββββ
β
ββββββΌββββββ
β SQLite β
β Message β
β Store β
ββββββββββββ
Messages flow: Agent β MCP call β gmsg server β SQLite β Recipient agent inbox
Message Schema
{
"id": "a0707298b3f6",
"sender": "EE",
"receiver": "Brand_Manager",
"message_type": "chat",
"message": "Full message content here",
"timestamp": "2026-05-10T16:20:28",
"read": false,
"conference_tag": null,
"archived": false
}
What Makes It Different
vs. Shared File Systems:
- Structured protocols - Not just dumping JSON files
- Delivery guarantees - Messages persist until read
- Conference support - Multi-party coordination built-in
vs. REST APIs:
- Async by design - No polling, messages persist until agent starts
- Agent-native - MCP integration, not HTTP endpoints
- Offline-capable - Works when recipient isnβt running
vs. Message Queues (RabbitMQ, etc.):
- Simple SQLite backend - No infrastructure overhead
- Claude Code optimized - Designed for AI agent communication patterns
- Read/unread semantics - Agents know whatβs new, not just FIFO
Tech Stack
- Python - Core message broker
- SQLite - Message storage and querying
- MCP (Model Context Protocol) - Claude Code integration
- JSON - Message serialization
Status
Production β Actively used by EE, Brand_Manager, IATV, MM, and other Silver Wizard Software agents.
Internal Infrastructure: gmsg is production infrastructure developed for Silver Wizard Softwareβs multi-agent coordination. Available for licensing to organizations building agentic systems.
MCP Tools Available
mcp__gmsg__send_message - Send a message
mcp__gmsg__check_messages - Read inbox
mcp__gmsg__get_message - Get specific message by ID
mcp__gmsg__mark_message_read - Mark as read
mcp__gmsg__archive_message - Archive old message
mcp__gmsg__delete_message - Delete message
mcp__gmsg__forward_message - Forward to another agent
mcp__gmsg__tag_conference - Tag message with conference
mcp__gmsg__untag_conference - Remove conference tag
Best Practices
- Use descriptive senders - βEEβ not βagent1β
- Conference tag liberally - Easier to filter later
- Archive, donβt delete - Preserve message history for debugging
- Include context in messages - Recipients may not have full conversation history
- Check inbox at session start - Use startup hooks to auto-check messages
Integration with CONF & IATV
CONF: Conference framework uses gmsg for agent coordination. Agents in a conference send/receive messages through gmsgβs conference tagging.
IATV: Auto-spawns agents when they receive messages. gmsg notifies IATV of unread messages, IATV launches the recipient agent.
CC_Mem: Agents often save message summaries to persistent memory for long-term recall.
Licensing
gmsg is internal production infrastructure available for licensing to organizations building multi-agent AI systems.
Contact: silver.wizard001@proton.me for licensing inquiries.
Documentation
- Getting Started Guide
- Message Protocol Specification
- Conference Tagging
- MCP Integration
- Message Lifecycle
Support
- GitHub: https://github.com/SilverWizardAI/gmsg (in LIB repo)
- Email: support@silverwizard.dev
- Licensing: silver.wizard001@proton.me