shared-context-server 1.1.10

Shared Context MCP Server for multi-agent collaboration

Shared Context Server

Content Navigation

Symbol	Meaning	Time Investment
🚀	Quick start	2-5 minutes
⚙️	Configuration	10-15 minutes
🔧	Deep dive	30+ minutes
💡	Why this works	Context only
⚠️	Important note	Read carefully

---

🎯 Quick Understanding (30 seconds)

A shared workspace for AI agents to collaborate on complex tasks.

The Problem: AI agents work independently, duplicate research, and can't build on each other's discoveries.

The Solution: Shared sessions where agents see previous findings and build incrementally instead of starting over.

# Agent 1: Security analysis
session.add_message("security_agent", "Found SQL injection in user login")

# Agent 2: Performance review (sees security findings)
session.add_message("perf_agent", "Optimized query while fixing SQL injection")

# Agent 3: Documentation (has full context)
session.add_message("docs_agent", "Documented secure, optimized login implementation")

Each agent builds on previous work instead of starting over.

💡 Uses MCP Protocol: Model Context Protocol - the standard for AI agent communication (works with Claude Code, Gemini, VS Code, Cursor, and frameworks like CrewAI).

---

🚀 Try It Now (2 minutes)

⚠️ Important: Choose Your Deployment Method

Docker (Recommended for Multi-Client Collaboration):

• ✅ Shared context across all MCP clients (Claude Code + Cursor + Windsurf)
• ✅ Persistent service - single server instance on port 23456
• ✅ True multi-agent collaboration - agents share sessions and memory
• 🎯 Use when: You want multiple tools to collaborate on the same tasks

uvx (Quick Trial & Testing Only):

• ⚠️ Isolated per-client - each MCP client gets its own separate instance
• ⚠️ No shared context - Claude Code and Cursor can't see each other's work
• ✅ Quick testing - perfect for trying features without Docker setup
• 🎯 Use when: Quick feature testing or learning the MCP tools in isolation

# 🐳 Docker: Multi-client shared collaboration (RECOMMENDED)
# ⚠️ Requires environment variables - see Step 1 below

# 📦 uvx: Isolated single-client testing only
# ⚠️ Requires API key - see Step 1 below
uvx shared-context-server --help

💡 TL;DR: Use Docker for real multi-agent work, uvx for quick testing only.

Prerequisites Check (30 seconds)

Choose your path:

• ✅ Docker (recommended): docker --version works
• ✅ uvx Trial: uvx --version works (testing only)

Environment Configuration Templates

Choose your .env template (for local development):

# 🚀 Quick Start (recommended) - Essential variables only
cp .env.minimal .env

# 🔧 Full Development - All development features
cp .env.example .env

# 🐳 Docker Deployment - Container-optimized paths
cp .env.docker .env

💡 Most users want .env.minimal - it contains only the 12 essential variables you actually need.

Step 1: Generate Keys & Start Server

🚀 One-Command Setup (Recommended)

# Clone and generate everything automatically
git clone https://github.com/leoric-crown/shared-context-server.git
cd shared-context-server
python scripts/generate_keys.py --docker-only
# ↳ Generates keys, shows Docker commands, creates .env file

Option A: Docker Compose (Recommended)

# After running the key generator above, choose your deployment:

# 🚀 Production (pre-built image from GHCR):
make docker
# OR: docker compose up -d

# 🔧 Development (with hot reload):
make dev-docker
# OR: docker compose -f docker-compose.dev.yml up -d

# 🏗️ Production (build locally):
make docker-local
# OR: docker compose -f docker-compose.yml -f docker-compose.local.yml up -d

Alternative: Raw Docker Commands

# If you prefer docker run over docker compose:
docker run -d --name shared-context-server -p 23456:23456 \
  -e API_KEY="your-generated-api-key" \
  -e JWT_SECRET_KEY="your-generated-jwt-secret" \
  -e JWT_ENCRYPTION_KEY="your-generated-jwt-encryption-key" \
  ghcr.io/leoric-crown/shared-context-server:latest

Option B: uvx Trial (Isolated Testing Only)

# Generate keys first
python scripts/generate_keys.py --uvx --no-file
# ↳ Shows the exact uvx command with generated keys

# Example output command to run:
API_KEY="generated-key" JWT_SECRET_KEY="generated-secret" \
  uvx shared-context-server --transport http

# ⚠️ IMPORTANT: Each MCP client gets isolated instances
# No shared context between Claude Code, Cursor, Windsurf

Option C: Local Development

# Full development setup
git clone https://github.com/leoric-crown/shared-context-server.git
cd shared-context-server
uv sync
python scripts/generate_keys.py --local
# ↳ Creates .env file and shows make dev command

make dev  # Starts with hot reload

Step 2: Connect Your MCP Client

The key generation script shows the exact commands with your API key. Replace YOUR_API_KEY_HERE with your generated key:

# Claude Code (simple HTTP transport)
claude mcp add --transport http scs http://localhost:23456/mcp/ \
  --header "X-API-Key: YOUR_API_KEY_HERE"

# Gemini CLI
gemini mcp add scs http://localhost:23456/mcp -t http -H "X-API-Key: YOUR_API_KEY_HERE"

# Test connection
claude mcp list  # Should show: ✓ Connected

VS Code Configuration

Add to your existing .vscode/mcp.json (create if it doesn't exist):

{
  "servers": {
    "shared-context-server": {
      "type": "http",
      "url": "http://localhost:23456/mcp",
      "headers": {"X-API-Key": "YOUR_API_KEY_HERE"}
    }
  }
}

Cursor Configuration

Add to your existing .cursor/mcp.json (create if it doesn't exist):

{
  "mcpServers": {
    "shared-context-server": {
      "command": "mcp-proxy",
      "args": ["--transport=streamablehttp", "http://localhost:23456/mcp/", "--headers", "X-API-Key", "YOUR_API_KEY_HERE"]
    }
  }
}

Claude Desktop Configuration

Add to your existing claude_desktop_config.json:

On MacOS, you may have to provide explicity path to mcp-proxy.

Have not tested in Windows.

{
    "scs": {
      "command": "/Users/YOUR_USER/.local/bin/mcp-proxy",
      "args": ["--transport=streamablehttp", "http://localhost:23456/mcp/", "--headers", "X-API-Key", "YOUR_API_KEY_HERE"]
    }
}

Step 3: Verify & Monitor

📝 Note: If you used make docker-prod, press Ctrl+C to exit the log viewer first, then run these commands in the same terminal.

# Test your setup (30 seconds)
# Method 1: Quick health check
curl http://localhost:23456/health

# Method 2: Create actual test session (see it in web UI!)
# If you have Claude Code with shared-context-server MCP tools:
# Run this in Claude: Create a session with purpose "README test setup"
# Expected: {"success": true, "session_id": "session_...", ...}

# Method 3: Test MCP tools discovery
npx @modelcontextprotocol/inspector --cli --method tools/list \
  -e API_KEY=$API_KEY \
  -e JWT_SECRET_KEY=$JWT_SECRET_KEY \
  -e JWT_ENCRYPTION_KEY=$JWT_ENCRYPTION_KEY \
  uv run python -m shared_context_server.scripts.cli

# Expected: {"tools": [...]} (proves MCP tools are available)

# Method 4: For Docker deployment, test via HTTP endpoint
npx @modelcontextprotocol/inspector --cli --method tools/list \
  http://localhost:23456/mcp

# View the dashboard
open http://localhost:23456/ui/  # Real-time session monitoring

✅ Success indicators:

• Health endpoint returns {"status": "healthy", ...}
• Dashboard loads at http://localhost:23456/ui/ and shows active sessions
• MCP Inspector validation error (proves MCP protocol is working)
• MCP client shows ✓ Connected status

📊 Web Dashboard (MVP)

Real-time monitoring interface for agent collaboration:

• Live session overview with active agent counts
• Real-time message streaming without page refreshes
• Session isolation visualization to track multi-agent workflows
• Performance monitoring for collaboration efficiency

💡 Perfect for: Monitoring agent handoffs, debugging collaboration flows, and demonstrating multi-agent coordination to stakeholders.

📦 PyPI Installation (Alternative Method)

The shared-context-server is also available on PyPI for quick testing:

# 📦 Install and try (creates isolated instances per client)
uvx shared-context-server --help
uvx shared-context-server --version

# ⚠️ For multi-client collaboration, use Docker instead

💡 When to use PyPI/uvx: Quick feature testing, learning MCP tools, single-client workflows only.

---

🔧 Choose Your Path

Are you...

├── 👨‍💻 Building a side project?
│   → [Simple Integration](#-simple-integration) (5 minutes)
│
├── 🏢 Planning enterprise deployment?
│   → [Enterprise Setup](#-enterprise-considerations) (15+ minutes)
│
├── 🎓 Researching multi-agent systems?
│   → [Technical Deep Dive](#-technical-architecture) (30+ minutes)
│
└── 🤔 Just evaluating the concept?
    → [Framework Integration Examples](#-framework-examples) (5 minutes)

---

🚀 Simple Integration

Works with existing tools you already use:

Direct MCP Integration (Tested)

# Via Claude Code or any MCP client
claude mcp add-json shared-context-server '{"command": "mcp-proxy", "args": ["--transport=streamablehttp", "http://localhost:23456/mcp/"]}'

# Direct MCP usage (use proper MCP client in production)
# Example shows concept - use mcp-proxy or MCP client libraries
import asyncio
from mcp_client import MCPClient  # Conceptual - use actual MCP client

async def create_session():
    client = MCPClient("http://localhost:23456/mcp/")
    return await client.call_tool("create_session", {"purpose": "agent collaboration"})

⚠️ Framework Integration Status: Direct MCP protocol tested. CrewAI, AutoGen, and LangChain integrations are conceptual - we welcome community contributions to develop and test these patterns.

➡️ Next: MCP Integration Examples

---

⚙️ Framework Examples

Code Review Pipeline

1. Security Agent finds vulnerabilities → shares findings
2. Performance Agent builds on security context → optimizes safely
3. Documentation Agent documents complete solution

💡 Why this works: Each agent builds on discoveries instead of duplicating work.

Research & Implementation

1. Research Agent gathers requirements → shares insights
2. Architecture Agent designs using research → documents decisions
3. Developer Agent implements with full context

More examples: Collaborative Workflows Guide

What works: ✅ MCP clients (Claude Code, Gemini, VS Code, Cursor) What's conceptual: 🔄 Framework patterns (CrewAI, AutoGen, LangChain) - community contributions welcome

---

🔧 What This Is / What This Isn't

✅ What this MCP server provides

• Real-time collaboration substrate for multi-agent workflows
• Session isolation with clean boundaries between different tasks
• MCP protocol compliance that works with any MCP-compatible agent framework
• Infrastructure layer that enhances existing orchestration tools

💡 Why MCP protocol? Universal compatibility - works with Claude Code, CrewAI, AutoGen, LangChain, and custom frameworks without vendor lock-in.

❌ What this MCP server isn't

• Not a vector database - Use Pinecone, Milvus, or Chroma for long-term storage
• Not an orchestration platform - Use CrewAI, AutoGen, or LangChain for task management
• Not for permanent memory - Sessions are for active collaboration, not archival

💡 Why this approach? We enhance your existing tools rather than replacing them - no need to rewrite your agent workflows.

---

🏢 Enterprise Considerations

⚙️ Production Setup & Scaling

Development → Production Path

Development (SQLite)

• ✅ Zero configuration
• ✅ Perfect for prototyping
• ❌ Limited to ~5 concurrent agents

Production (PostgreSQL)

• ✅ High concurrency (20+ agents)
• ✅ Enterprise backup/recovery
• ❌ Requires database management

Enterprise Features Roadmap

• SSO Integration: SAML/OIDC support planned
• Audit Logging: Enhanced compliance logging
• High Availability: Multi-node deployment
• Advanced RBAC: Attribute-based permissions

Migration: Start with SQLite, migrate when you hit concurrency limits.

🔧 Security & Compliance

Current Security Features

• JWT Authentication: Role-based access control
• Input Sanitization: XSS and injection prevention
• Secure Token Management: Prevents JWT exposure vulnerabilities
• Message Visibility: Public/private/agent-only filtering

Enterprise Security Roadmap

• SSO Integration: SAML, OIDC, Active Directory
• Audit Trails: SOX, HIPAA-compliant logging
• Data Governance: Retention policies, geographic residency
• Advanced Encryption: At-rest and in-transit encryption

---

🔧 Technical Architecture

🔄 Deployment Architecture: Docker vs uvx

Docker Deployment (Multi-Client Shared Context)

┌─────────────────┐    ┌──────────────────────┐
│   Claude Code   │───▶│                      │
├─────────────────┤    │  Shared HTTP Server  │
│     Cursor      │───▶│   (port 23456)       │
├─────────────────┤    │                      │
│    Windsurf     │───▶│  • Single database   │
└─────────────────┘    │  • Shared sessions   │
                       │  • Cross-tool memory │
                       └──────────────────────┘

✅ Enables: True multi-agent collaboration, session sharing, persistent context

uvx Deployment (Isolated Per-Client)

┌─────────────────┐    ┌─────────────────┐
│   Claude Code   │───▶│ Isolated Server │
└─────────────────┘    │ + Database #1   │
                       └─────────────────┘
┌─────────────────┐    ┌─────────────────┐
│     Cursor      │───▶│ Isolated Server │
└─────────────────┘    │ + Database #2   │
                       └─────────────────┘
┌─────────────────┐    ┌─────────────────┐
│    Windsurf     │───▶│ Isolated Server │
└─────────────────┘    │ + Database #3   │
                       └─────────────────┘

⚠️ Limitation: No cross-tool collaboration, separate contexts, testing only

💡 Key Insight: Docker provides the "shared" in shared-context-server, while uvx creates isolated silos.

Core Design Principles

Session-Based Isolation

What: Each collaborative task gets its own workspace Why: Prevents cross-contamination while enabling rich collaboration within teams

Message Visibility Controls

What: Four-tier system (public/private/agent-only/admin-only) Why: Granular information sharing - agents can have private working memory and shared discoveries

MCP Protocol Integration

What: Model Context Protocol compliance for universal compatibility Why: Works with any MCP-compatible framework without custom integration code

Performance Characteristics

Designed for Real-Time Collaboration

• <30ms message operations for smooth agent handoffs
• 2-3ms fuzzy search across session history
• 20+ concurrent agents per session
• Session continuity during agent switches

💡 Why these targets? Sub-30ms ensures imperceptible delays during agent handoffs, maintaining workflow momentum.

Scalability Considerations

• SQLite: Development and small teams (<5 concurrent agents)
• PostgreSQL: Production deployments (20+ concurrent agents)
• Connection pooling: Built-in performance optimization
• Multi-level caching: >70% cache hit ratio for common operations

Database & Storage

Architecture Decision: Database Choice

SQLite for Development

• ✅ Zero configuration
• ✅ Perfect for prototyping
• ❌ Single writer limitation

PostgreSQL for Production

• ✅ Multi-writer concurrency
• ✅ Enterprise backup/recovery
• ✅ Advanced indexing and performance
• ❌ Requires database administration

Database Backend

• Unified: SQLAlchemy Core (supports SQLite, PostgreSQL, MySQL)
• Development: SQLite with aiosqlite driver (fastest, simplest)
• Production: PostgreSQL/MySQL with async drivers (scalable, robust)

Migration Path: SQLAlchemy backend provides smooth transition to PostgreSQL when scaling needs arise.

💡 Why this hybrid approach? Optimizes for developer experience during development while supporting enterprise scale in production.

---

📖 Documentation & Next Steps

🟢 Getting Started Paths

• Integration Guide - CrewAI, AutoGen, LangChain examples
• Quick Reference - Commands and common tasks
• Development Setup - Local development environment

🟡 Production Deployment

• Docker Setup - Container deployment guide
• API Reference - All 15+ MCP tools with examples
• Troubleshooting - Common issues and solutions

🔴 Advanced Topics

• Custom Integration - Build your own MCP integration
• Production Deployment - Docker and scaling strategies

All documentation: Documentation Index

---

🚀 Development Commands

make help        # Show all available commands
make dev         # Start development server with hot reload
make test        # Run tests with coverage
make quality     # Run all quality checks
make docker      # Production Docker (GHCR image) → shows logs
make dev-docker  # Development Docker (local build + hot reload) → shows logs
# ⚠️ Both commands show live logs - press Ctrl+C to exit and continue setup

⚙️ Direct commands without make

# Development
uv sync && uv run python -m shared_context_server.scripts.dev

# Testing
uv run pytest --cov=src

# Quality checks
uv run ruff check && uv run mypy src/

---

License

MIT License - Open source software for the AI community.

---

Built with modern Python tooling and MCP standards. Contributions welcome!